cs.CVApr 9, 2026

⊘\oslash Source Models Leak What They Shouldn't ↛\nrightarrow: Unlearning Zero-Shot Transfer in Domain Adaptation Through Adversarial Optimization

Authors: Arnav Devalapally, Poornima Jain, Kartik Srinivas, Vineeth N. Balasubramanian

Organizations: Indian Institute of Technology, Hyderabad · University of Michigan · Carnegie Mellon University · Microsoft Research

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The increasing adaptation of vision models across domains, such as satellite imagery and medical scans, has raised an emerging privacy risk: models may inadvertently retain and leak sensitive source-domain specific information in the target domain. This creates a compelling use case for machine unlearning to protect the privacy of sensitive source-domain data. Among adaptation techniques, source-free domain adaptation (SFDA) calls for an urgent need for machine unlearning (MU), where the source data itself is protected, yet the source model exposed during adaptation encodes its influence. Our experiments reveal that existing SFDA methods exhibit strong zero-shot performance on source-exclusive classes in the target domain, indicating they inadvertently leak knowledge of these classes into the target domain, even when they are not represented in the target data. We identify and address this risk by proposing an MU setting called SCADA-UL: Unlearning Source-exclusive ClAsses in Domain Adaptation. Existing MU methods do not address this setting as they are not designed to handle data distribution shifts. We propose a new unlearning method, where an adversarially generated forget class sample is unlearned by the model during the domain adaptation process using a novel rescaled labeling strategy and adversarial optimization. We also extend our study to two variants: a continual version of this problem setting and to one where the specific source classes to be forgotten may be unknown. Alongside theoretical interpretations, our comprehensive empirical results show that our method consistently outperforms baselines in the proposed setting while achieving retraining-level unlearning performance on benchmark datasets. Our code is available at https://github.com/D-Arnav/SCADA

Figures & tables

Figure 1: Comparison of Conventional MU and Proposed SCADA-UL. Conventional class-wise machine unlearning focuses on forgetting a subset of classes from a trained model. On the other hand, the proposed SCADA-UL aims to remove knowledge of source-exclusive classes (classes absent in target domain) while adapting a model to a new domain. For instance, if a land-use categorization model is adapted to a new geography, sensitive classes such as airports must not be transferred to the target domain.
Figure 2: Existing SFDA/PDA methods leak source-exclusive classes in the target domain.
Existing EffortsUnlearningAdaptationForget Data-free
SFDA/PDA [ 34 , 28 , 5 ]✗✓N/A
MU [ 21 , 15 ]✓✗✗
Zero Glance MU [ 51 , 10 ]✓✗✓
SCADA-UL✓✓✓
Table 1: Comparison of our setting with existing efforts
Figure 3: Our overall method. We adapt a source trained model wSw^{\mathcal{S}} to a target domain DT\mathcal{D}^{\mathcal{T}} while forgetting classes cFc_{\mathcal{F}} . We first create an initial adversarial sample x^\hat{x} by maximizing its probability of belonging to the forget classes cFc_{\mathcal{F}} by optimizing LADV(wS,x^)\mathcal{L}_{\text{ADV}}(w^{\mathcal{S}},\hat{x}) (Eqn 3 ). In each subsequent iteration, the model wT\smash{w^{\mathcal{T}}} minimizes : (i) LMU\mathcal{L}_{\text{MU}} ( 5 ) using rescaled labels y^\hat{y} to encourage unlearning the forget classes cFc_{\mathcal{F}} ; and (ii) LSFDA\mathcal{L}_{\text{SFDA}} to simultaneously adapt to the target domain. Before the end of each iteration, the adversarial sample x^\hat{x} is re-optimized to maximize its probability of belonging to the forget classes by maximizing - LADV\mathcal{L}_{\text{ADV}} .
 OfficeHomeOffice31DomainNet-126
MethodADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrowScore ↑\uparrowADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrowScore ↑\uparrowADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrowScore ↑\uparrow
Original (SF(DA) 2 [ 28 ] )75.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}58.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}76.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}90.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}38.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Retrain76.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}77.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.77_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}66.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.66_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Finetune\textbf{76.1}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}49.2_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}\underline{76.7}_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}82.8_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}\underline{66.5}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}20.4_{{\color[rgb]{0.5,0.5,0.5}\pm 4.1}}\underline{0.56}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
UNSIR [ 51 ]35.0_{{\color[rgb]{0.5,0.5,0.5}\pm 5.6}}\textbf{0.0}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.35_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}59.7_{{\color[rgb]{0.5,0.5,0.5}\pm 5.0}}\underline{16.3}_{{\color[rgb]{0.5,0.5,0.5}\pm 3.3}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}14.6_{{\color[rgb]{0.5,0.5,0.5}\pm 3.6}}\underline{0.4}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}0.14_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
ZSMU [ 10 ]71.1_{{\color[rgb]{0.5,0.5,0.5}\pm 6.1}}39.8_{{\color[rgb]{0.5,0.5,0.5}\pm 10.}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}77.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}84.2_{{\color[rgb]{0.5,0.5,0.5}\pm 3.3}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}65.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}34.5_{{\color[rgb]{0.5,0.5,0.5}\pm 6.6}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table 2: Results for Multi-Class SCADA-UL on OfficeHome, Office31, DomainNet datasets. We compare against 10 baselines: Original = model adapted to target domain without applying unlearning; Retrain = model adapted with source model retrained without forget data; Finetune = model finetuned on subset of target retain data; MU methods [ 51 , 10 , 15 , 53 , 2 ] ; PDA/SFPDA methods [ 5 , 34 ] . ADfTA_{\mathcal{D}_{\smash{f}}^{\mathcal{T}}} = accuracy on target forget set; ADrTA_{\mathcal{D}_{r}^{\mathcal{T}}} = target retain accuracy; Score captures overall unlearning performance i.e. high ADrTA_{\mathcal{D}_{r}^{\mathcal{T}}} + low ADfTA_{\mathcal{D}_{\smash{f}}^{\mathcal{T}}} . (Best result in bold, second-best underlined)
Table 6
StageAcc DrT↑{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowAcc DfT↓{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
Before Adaptation75.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}52.0_{{\color[rgb]{0.5,0.5,0.5}\pm 5.3}}
During Adaptation75.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
After Adaptation58.2_{{\color[rgb]{0.5,0.5,0.5}\pm 7.1}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table 5: Study on stage of applying our algorithm
StrategyAcc DrT↑{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowAcc DfT↓{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
Rescaled{75.1_{\color[rgb]{0.5,0.5,0.5}{\pm 1.3}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
Uniform{70.5_{\color[rgb]{0.5,0.5,0.5}{\pm 3.6}}}{3.5_{\color[rgb]{0.5,0.5,0.5}{\pm 4.1}}}
Random{1.5_{\color[rgb]{0.5,0.5,0.5}{\pm 0.4}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
Table 6: Choice of relabeling strategy
 y=fwT(x^cF)y=f_{w^{\mathcal{T}}}(\hat{x}_{c_{\mathcal{F}}}) 
Algorithm 2 Adversarial Optimization for SCADA-UL (AO_SCADA_UL: Detailed Algorithm)
Figure 10
Figure A.4 : Visual appearance of Adversarial Samples. Top row shows examples of samples from the source forget class (Backpack) from OfficeHome Art domain, while the bottom row shows examples of adversarial samples of this class. Although these samples appear as random noise, they are confidently classified as a backpack by the model. Moreover, in additional studies ( Fig. A.5 ), we show that these samples match the forget class even on a representational level)
Compared FeaturesSCS_{C} Before AdaptationSCS_{C} After Adaptation
Adversarial - Source Forget0.3600.371
Adversarial - Source Retain0.3260.315
Adversarial - Target Forget0.3590.366
Adversarial - Target Retain0.3350.307
Table A.8 : Cosine similarity ( SCS_{C} ) between features of adversarial samples and real samples. Adversarial sample features align more with those of the forget class both before and after adaptation. After adaptation, they align more with the target forget
Figure A.5 : t-SNE plots of adversarial, retain, and forget class samples. Left: Evolution of an adversarial sample over time. Initially, the sample lies far from the forget-class centroids (closer to retain-class centroids). While being optimized on the source model, it moves closer to the source forget-class centroid, and over iterations gradually converges toward the target forget-class centroid. This shows that the samples evolve alongside the model to best fit the class to forget in the target domain. Right: Final adversarial samples compared to t-SNE embeddings of 9 randomly selected retain classes {29,3,44,58,36,2,59,47,14}\{29,3,44,58,36,2,59,47,14\} and the forget class cF=1c_{\mathcal{F}}=1 . The final adversarial samples clearly align more closely with features from the forget class. Plotted for OfficeHome Art to Product.
Figure A.6 : MU Loss ( LMU\mathcal{L}_{\text{MU}} ) and Adversarial Loss ( LADV\mathcal{L_{\text{ADV}}} ). Initially, LMU\mathcal{L}_{\text{MU}} conflicts with LADV\mathcal{L}_{\text{ADV}} , but equilibrium is reached after ∼100\sim 100 steps. This demonstrates a measured approach to unlearning where the model gradually finds a new representation ϕ\phi that aligns with both objectives.
Figure A.7 : Softmax outputs of the target forget set for the retrained model compared with our method using Rescaled, Uniform, and Random Labeling. After sorting classes by the retrained model’s outputs, our method with Rescaled Labeling best matches the retrained distribution, while Uniform and Random Labeling deviate more significantly, as also reflected in their higher KL Divergence values. Note that we use interpolation to make the distributions more clear.
InitializeADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
✓{75.1_{\color[rgb]{0.5,0.5,0.5}{\pm 1.3}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
✗{75.0_{\color[rgb]{0.5,0.5,0.5}{\pm 1.2}}}{15.2_{\color[rgb]{0.5,0.5,0.5}{\pm 4.1}}}
Table A.11 : Initializing adv samples
α\alphaADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
1.01.0{75.0_{\color[rgb]{0.5,0.5,0.5}{\pm 1.2}}}{15.2_{\color[rgb]{0.5,0.5,0.5}{\pm 4.1}}}
5.05.0{75.2_{\color[rgb]{0.5,0.5,0.5}{\pm 1.1}}}{2.6_{\color[rgb]{0.5,0.5,0.5}{\pm 2.0}}}
10.010.0{75.1_{\color[rgb]{0.5,0.5,0.5}{\pm 1.3}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
20.020.0{75.2_{\color[rgb]{0.5,0.5,0.5}{\pm 1.1}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
Table A.12 : Effect of varying Loss Trade-off ( α\alpha ) hyperparameter study
NadvN_{\text{adv}}ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
11{70.9_{\color[rgb]{0.5,0.5,0.5}{\pm 3.6}}}{0.8_{\color[rgb]{0.5,0.5,0.5}{\pm 1.4}}}
22{72.9_{\color[rgb]{0.5,0.5,0.5}{\pm 3.7}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
44{75.1_{\color[rgb]{0.5,0.5,0.5}{\pm 1.3}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
88{75.2_{\color[rgb]{0.5,0.5,0.5}{\pm 1.2}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
1616{75.3_{\color[rgb]{0.5,0.5,0.5}{\pm 1.1}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
Table A.13 : Study on # adversarial samples used during training
EpochsADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrowADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow
11{75.6_{\color[rgb]{0.5,0.5,0.5}{\pm 1.4}}}{18.2_{\color[rgb]{0.5,0.5,0.5}{\pm 8.1}}}
55{75.1_{\color[rgb]{0.5,0.5,0.5}{\pm 1.3}}}{0.0_{\color[rgb]{0.5,0.5,0.5}{\pm 0.0}}}
1010{74.9_{\color[rgb]{0.5,0.5,0.5}{\pm 1.2}}}{0.4_{\color[rgb]{0.5,0.5,0.5}{\pm 0.3}}}
Table A.14 : Study on number of train epochs
SFDA LossMethodMetrics →\rightarrow pc →\rightarrow sp →\rightarrow cp →\rightarrow rr →\rightarrow sr →\rightarrow cr →\rightarrow pAverage% Improved
SHOT [ 34 ]OriginalADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow70.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}57.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}61.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}83.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}60.6_{{\color[rgb]{0.5,0.5,0.5}\pm 3.4}}69.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}75.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}68.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}} 
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow88.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}47.0_{{\color[rgb]{0.5,0.5,0.5}\pm 20.6}}57.8_{{\color[rgb]{0.5,0.5,0.5}\pm 2.5}}86.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}27.7_{{\color[rgb]{0.5,0.5,0.5}\pm 4.8}}41.8_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}43.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}56.1_{{\color[rgb]{0.5,0.5,0.5}\pm 4.5}}-
Score ↑\uparrow0.37_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.39_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.39_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}} 
Our MethodADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow56.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}36.6_{{\color[rgb]{0.5,0.5,0.5}\pm 4.0}}42.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}77.9_{{\color[rgb]{0.5,0.5,0.5}\pm 5.5}}27.1_{{\color[rgb]{0.5,0.5,0.5}\pm 5.3}}49.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}66.4_{{\color[rgb]{0.5,0.5,0.5}\pm 5.2}}50.9_{{\color[rgb]{0.5,0.5,0.5}\pm 3.2}} 
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}-
Score ↑\uparrow\underline{0.56}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.37_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}\underline{0.78}_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}0.27_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.66_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}} 
Table A.17 : Improved SFDA methods improve our method as well. We observe that as the underlying SFDA method achieves better performance, our method also improves. This shows the adaptability of our approach to new SFDA loss terms and its potential to further benefit as future SFDA methods advance.
Figure A.8 : Forget/Retain Error and Calibration Analysis. The figure shows the False Negative Rate (Forget FNR) of our method on forget-class samples across thresholds (first column); the False Positive Rate (Retain FPR) on retain-class samples across thresholds (second column); Forget FNR vs. Retain FPR (third column); and calibration error on retain and forget samples before and after unlearning (fourth column) for two OfficeHome settings. Our method achieves high Forget FNR (the model is confused on forget samples) and low Retain FPR (retain samples remain correctly classified) at appropriate thresholds. Calibration error is high for both forget and retain classes on the source model; after unlearning, retain-class calibration error stays low while forget-class calibration error increases, indicating greater model uncertainty on forget samples.
MethodMetricA →\rightarrow CA →\rightarrow PA →\rightarrow RC →\rightarrow AC →\rightarrow PC →\rightarrow RP →\rightarrow AP →\rightarrow CP →\rightarrow RR →\rightarrow AR →\rightarrow CR →\rightarrow PAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow61.0_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}82.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}82.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}76.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.1}}82.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}82.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}75.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.7}}58.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}84.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}58.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}87.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}75.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow28.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}71.1_{{\color[rgb]{0.5,0.5,0.5}\pm 6.2}}86.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}57.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}69.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}76.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}47.8_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}22.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}72.8_{{\color[rgb]{0.5,0.5,0.5}\pm 6.9}}53.1_{{\color[rgb]{0.5,0.5,0.5}\pm 4.2}}35.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}77.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}58.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}
Score ↑\uparrow0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow60.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}83.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}82.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.5}}82.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}83.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}76.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}59.7_{{\color[rgb]{0.5,0.5,0.5}\pm 3.1}}84.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}77.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}59.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}87.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}76.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.82_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.79_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.84_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.87_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.18 : Results for Multi-Class SCADA Unlearning on OfficeHome. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodMetrics →\rightarrow pc →\rightarrow sp →\rightarrow cp →\rightarrow rr →\rightarrow sr →\rightarrow cr →\rightarrow pAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}66.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}61.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}78.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}55.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}65.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}75.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow67.6_{{\color[rgb]{0.5,0.5,0.5}\pm 7.0}}55.9_{{\color[rgb]{0.5,0.5,0.5}\pm 5.9}}42.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.7}}60.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.7}}6.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}14.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}22.8_{{\color[rgb]{0.5,0.5,0.5}\pm 3.8}}38.7_{{\color[rgb]{0.5,0.5,0.5}\pm 3.5}}
Score ↑\uparrow0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.57_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}65.3_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}58.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}54.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}62.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}74.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}66.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.71_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.65_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.59_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.54_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.63_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.74_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.66_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.19 : Results for Multi-Class SCADA Unlearning on DomainNet. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodMetricA →\rightarrow DA →\rightarrow WD →\rightarrow AD →\rightarrow WW →\rightarrow AW →\rightarrow DAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow77.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}86.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}67.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}79.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}72.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}76.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow84.8_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}99.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}93.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}88.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}97.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}76.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}90.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}
Score ↑\uparrow0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.35_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.36_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow76.0_{{\color[rgb]{0.5,0.5,0.5}\pm 5.0}}87.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}68.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}83.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}72.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}78.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}77.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}0.87_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.68_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.72_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.77_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.20 : Results for Multi-Class SCADA Unlearning on Office 31. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodMetricA →\rightarrow CA →\rightarrow PA →\rightarrow RC →\rightarrow AC →\rightarrow PC →\rightarrow RP →\rightarrow AP →\rightarrow CP →\rightarrow RR →\rightarrow AR →\rightarrow CR →\rightarrow PAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow60.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}82.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}85.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}74.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}83.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}85.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}72.1_{{\color[rgb]{0.5,0.5,0.5}\pm 3.3}}59.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}86.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}75.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}60.0_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}88.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}76.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow51.2_{{\color[rgb]{0.5,0.5,0.5}\pm 9.8}}88.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}89.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}42.3_{{\color[rgb]{0.5,0.5,0.5}\pm 7.5}}72.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.7}}85.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}54.5_{{\color[rgb]{0.5,0.5,0.5}\pm 8.6}}50.0_{{\color[rgb]{0.5,0.5,0.5}\pm 6.4}}88.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}39.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}80.4_{{\color[rgb]{0.5,0.5,0.5}\pm 3.6}}93.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}69.7_{{\color[rgb]{0.5,0.5,0.5}\pm 4.2}}
Score ↑\uparrow0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.46_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.46_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.54_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.33_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow62.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}83.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}86.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}74.7_{{\color[rgb]{0.5,0.5,0.5}\pm 3.2}}82.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}86.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}72.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}60.0_{{\color[rgb]{0.5,0.5,0.5}\pm 3.1}}86.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}75.0_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}61.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}88.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}76.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.62_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.86_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.75_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.86_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.72_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.87_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.75_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.88_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.77_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.21 : Results for Single-Class SCADA Unlearning on OfficeHome. Forget class is cF=1c_{\mathcal{F}}=1 (Best result in bold, second-best underlined)
MethodMetrics →\rightarrow pc →\rightarrow sp →\rightarrow cp →\rightarrow rr →\rightarrow sr →\rightarrow cr →\rightarrow pAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}65.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}62.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}55.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.6}}65.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}75.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}67.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow77.6_{{\color[rgb]{0.5,0.5,0.5}\pm 5.7}}36.9_{{\color[rgb]{0.5,0.5,0.5}\pm 6.9}}35.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}74.1_{{\color[rgb]{0.5,0.5,0.5}\pm 12.1}}3.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}17.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}6.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}35.9_{{\color[rgb]{0.5,0.5,0.5}\pm 4.0}}
Score ↑\uparrow0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.46_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.54_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.55_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.71_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}66.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}61.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}78.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}56.0_{{\color[rgb]{0.5,0.5,0.5}\pm 3.2}}65.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}74.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}67.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.72_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.67_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.56_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.65_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.74_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.68_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.22 : Results for Single-Class SCADA Unlearning on DomainNet. Forget class is cF=1c_{\mathcal{F}}=1 (Best result in bold, second-best underlined)
MethodMetricA →\rightarrow DA →\rightarrow WD →\rightarrow AD →\rightarrow WW →\rightarrow AW →\rightarrow DAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow78.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}81.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}61.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}76.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}67.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}80.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}74.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow91.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}100._{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}100._{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}100._{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}100._{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}90.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}97.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}
Score ↑\uparrow0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.31_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.38_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.34_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.38_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow79.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}82.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}63.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.1}}67.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}78.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}74.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.79_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.82_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.64_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.67_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.75_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.23 : Results for Single-Class SCADA Unlearning on Office 31. Forget class is cF=1c_{\mathcal{F}}=1 (Best result in bold, second-best underlined)
MethodMetricA →\rightarrow CA →\rightarrow PA →\rightarrow RC →\rightarrow AC →\rightarrow PC →\rightarrow RP →\rightarrow AP →\rightarrow CP →\rightarrow RR →\rightarrow AR →\rightarrow CR →\rightarrow PAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow61.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}82.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}82.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}76.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.7}}82.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}82.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}75.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}58.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}84.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}58.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}87.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}75.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow28.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}71.1_{{\color[rgb]{0.5,0.5,0.5}\pm 5.0}}86.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}57.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}69.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}76.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}47.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}22.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}72.8_{{\color[rgb]{0.5,0.5,0.5}\pm 5.7}}53.1_{{\color[rgb]{0.5,0.5,0.5}\pm 3.4}}35.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}77.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.7}}58.1_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}
Score ↑\uparrow0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow60.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}83.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}82.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.1}}82.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}83.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}76.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}59.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.5}}84.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}77.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}59.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}87.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}76.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.82_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.79_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.84_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.87_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.24 : Results for UC-SCADA Unlearning on OfficeHome. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodMetrics →\rightarrow pc →\rightarrow sp →\rightarrow cp →\rightarrow rr →\rightarrow sr →\rightarrow cr →\rightarrow pAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}66.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}61.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}78.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}55.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}65.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}75.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow67.6_{{\color[rgb]{0.5,0.5,0.5}\pm 5.7}}55.9_{{\color[rgb]{0.5,0.5,0.5}\pm 4.8}}42.9_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}60.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.2}}6.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}14.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}22.8_{{\color[rgb]{0.5,0.5,0.5}\pm 3.1}}38.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}
Score ↑\uparrow0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.57_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.61_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}65.3_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}58.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}54.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}62.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}74.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}66.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.71_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.65_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.59_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.54_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.63_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.74_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.66_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.25 : Results for UC-SCADA Unlearning on DomainNet. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodMetricA →\rightarrow DA →\rightarrow WD →\rightarrow AD →\rightarrow WW →\rightarrow AW →\rightarrow DAverage
Original (SF(DA) 2 [ 28 ] )ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow67.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}79.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}77.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}86.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}78.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}72.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}76.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow93.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}88.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}84.8_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}99.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}76.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}97.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}90.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}
Score ↑\uparrow0.35_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.36_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow68.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}83.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}76.0_{{\color[rgb]{0.5,0.5,0.5}\pm 4.1}}87.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}72.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}77.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Score ↑\uparrow0.68_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.83_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.76_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.87_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.78_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.72_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.77_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.26 : Results for UC-SCADA Unlearning on Office 31. Forget classes are CF={1,2,3}\mathcal{C_{F}}=\{1,2,3\} (Best result in bold, second-best underlined)
MethodTaskAcc.A →\rightarrow CA →\rightarrow PA →\rightarrow RC →\rightarrow AC →\rightarrow PC →\rightarrow RP →\rightarrow AP →\rightarrow CP →\rightarrow RR →\rightarrow AR →\rightarrow CR →\rightarrow PAverage
Original (SF(DA) 2 [ 28 ] )T1ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow63.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}80.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}86.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}78.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}81.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}87.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}76.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}62.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}89.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}81.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}63.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}87.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow43.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}61.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}75.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}59.8_{{\color[rgb]{0.5,0.5,0.5}\pm 3.7}}55.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}66.9_{{\color[rgb]{0.5,0.5,0.5}\pm 3.2}}46.0_{{\color[rgb]{0.5,0.5,0.5}\pm 2.6}}26.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}62.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}58.4_{{\color[rgb]{0.5,0.5,0.5}\pm 4.8}}40.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}63.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}55.0_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}
Score ↑\uparrow0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.55_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
T2ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow63.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}80.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}86.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}78.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}81.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}87.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}76.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}62.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.3}}89.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}81.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}63.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}87.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow44.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}62.2_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}76.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}60.6_{{\color[rgb]{0.5,0.5,0.5}\pm 3.7}}56.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}67.5_{{\color[rgb]{0.5,0.5,0.5}\pm 3.2}}46.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.6}}26.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}62.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}59.2_{{\color[rgb]{0.5,0.5,0.5}\pm 4.9}}40.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}64.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}55.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}
Score ↑\uparrow0.44_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.49_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.55_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.53_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.50_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.27 : C-SCADA Unlearning performance on the OfficeHome dataset. Forget classes are CF1={1,2},CF2={3,4},CF3={5,6}\mathcal{C_{F}}^{1}=\{1,2\},\mathcal{C_{F}}^{2}=\{3,4\},\mathcal{C_{F}}^{3}=\{5,6\}
MethodTaskAcc.s →\rightarrow pc →\rightarrow sp →\rightarrow cp →\rightarrow rr →\rightarrow sr →\rightarrow cr →\rightarrow pAverage
Original (SF(DA) 2 [ 28 ] )T1ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.8_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}64.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}62.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}55.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}65.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}75.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow37.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}37.2_{{\color[rgb]{0.5,0.5,0.5}\pm 2.9}}37.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}63.2_{{\color[rgb]{0.5,0.5,0.5}\pm 3.6}}17.3_{{\color[rgb]{0.5,0.5,0.5}\pm 2.1}}10.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}25.5_{{\color[rgb]{0.5,0.5,0.5}\pm 3.5}}32.5_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}
Score ↑\uparrow0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
T2ADrT↑A_{\mathcal{D}_{r}^{\mathcal{T}}}\uparrow71.8_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}64.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}62.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.1}}55.6_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}65.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}75.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}
ADfT↓A_{\mathcal{D}_{f}^{\mathcal{T}}}\downarrow37.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}37.4_{{\color[rgb]{0.5,0.5,0.5}\pm 3.0}}37.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}63.3_{{\color[rgb]{0.5,0.5,0.5}\pm 3.6}}17.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.1}}10.1_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}25.7_{{\color[rgb]{0.5,0.5,0.5}\pm 3.5}}32.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.3}}
Score ↑\uparrow0.52_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.45_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.48_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.47_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.60_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.51_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
Table A.28 : C-SCADA Unlearning performance on the DomainNet dataset. Forget classes are CF1={1,2},CF2={3,4},CF3={5,6}\mathcal{C_{F}}^{1}=\{1,2\},\mathcal{C_{F}}^{2}=\{3,4\},\mathcal{C_{F}}^{3}=\{5,6\}
MethodMetricA →\rightarrow CA →\rightarrow PA →\rightarrow RC →\rightarrow AC →\rightarrow PC →\rightarrow RP →\rightarrow AP →\rightarrow CP →\rightarrow RR →\rightarrow AR →\rightarrow CR →\rightarrow PAverage
Original (SHOT [ 34 ] )OS Acc. ↑\uparrow54.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}69.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}75.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}70.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}72.5_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}76.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}67.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.6}}49.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}75.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}72.8_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}55.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}76.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}68.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}
OS* Acc. ↑\uparrow60.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}77.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}81.8_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}75.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}79.9_{{\color[rgb]{0.5,0.5,0.5}\pm 0.6}}83.0_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}72.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}54.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.9}}81.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}78.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.7}}61.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}84.3_{{\color[rgb]{0.5,0.5,0.5}\pm 1.4}}74.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}
Forget Acc. ↓\downarrow43.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}85.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}86.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.7}}68.2_{{\color[rgb]{0.5,0.5,0.5}\pm 2.7}}78.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}79.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}65.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}36.5_{{\color[rgb]{0.5,0.5,0.5}\pm 4.0}}83.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}70.7_{{\color[rgb]{0.5,0.5,0.5}\pm 1.1}}45.0_{{\color[rgb]{0.5,0.5,0.5}\pm 3.3}}89.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}69.4_{{\color[rgb]{0.5,0.5,0.5}\pm 1.6}}
Score ↑\uparrow0.38_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.38_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.42_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.36_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.41_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.43_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.38_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}0.40_{{\color[rgb]{0.5,0.5,0.5}\pm 0.0}}
RetrainOS Acc. ↑\uparrow51.7_{{\color[rgb]{0.5,0.5,0.5}\pm 5.5}}67.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}75.2_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}70.3_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}71.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}76.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.8}}70.7_{{\color[rgb]{0.5,0.5,0.5}\pm 2.8}}51.5_{{\color[rgb]{0.5,0.5,0.5}\pm 3.7}}77.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.4}}74.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}58.0_{{\color[rgb]{0.5,0.5,0.5}\pm 1.5}}77.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.4}}68.5_{{\color[rgb]{0.5,0.5,0.5}\pm 1.8}}
OS* Acc. ↑\uparrow57.2_{{\color[rgb]{0.5,0.5,0.5}\pm 6.1}}74.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}81.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.2}}75.6_{{\color[rgb]{0.5,0.5,0.5}\pm 1.0}}78.9_{{\color[rgb]{0.5,0.5,0.5}\pm 1.3}}82.4_{{\color[rgb]{0.5,0.5,0.5}\pm 0.9}}75.9_{{\color[rgb]{0.5,0.5,0.5}\pm 3.1}}56.9_{{\color[rgb]{0.5,0.5,0.5}\pm 4.1}}83.2_{{\color[rgb]{0.5,0.5,0.5}\pm 0.5}}79.7_{{\color[rgb]{0.5,0.5,0.5}\pm 0.2}}64.1_{{\color[rgb]{0.5,0.5,0.5}\pm 1.7}}85.4_{{\color[rgb]{0.5,0.5,0.5}\pm 2.7}}74.6_{{\color[rgb]{0.5,0.5,0.5}\pm 2.0}}
Table A.30 : Results for Multi-Class SCADA Unlearning on Openset Domain Adaptation on OfficeHome. Implemented with SHOT-ODA as the open-set domain adaptation loss
More in this topic

Machine Unlearning

Machine unlearning removes the influence of selected data or concepts from trained models while preserving useful behavior and avoiding full retraining.

See all 351 papers

Explore similar work

CardsList
May 4, 2026

Rethinking the Need for Source Models: Source-Free Domain Adaptation from Scratch Guided by a Vision-Language Model

Source-Free Domain Adaptation (SFDA) adapts source models to target domains without accessing source data, addressing privacy and transmission issues. However, existing methods still initialize from a source pre-trained model and thus are not truly source-free. Recent works have introduced Vision-Language (ViL) models to guide the adaptation process, in these methods, we observe that for the same target domain, different source models yield minimal variation in final results, indicating the source model itself has limited impact. Motivated by this, we propose ViL-Only Domain Adaptation (VODA) , a stricter setting that eliminates all dependencies on source domain, relying solely on a randomly initialized model, a ViL model, and unlabeled target data. We analyze the adaptation dynamics of VODA and introduce Two-Stage Denoised-Region Distillation (TS-DRD) , a two-stage framework that first warms up the model with ViL guidance, then seek a Denoised-Region inherent in both the ViL and adapting model, yielding cleaner supervision for distillation. Experiments on Office-Home, VisDA, and DomainNet-126 show that under VODA, TS-DRD achieves competitive or superior performance to existing SFDA methods that still use source models, demonstrating its effectiveness and the potential of the VODA setting.

Zhou Bingtao, Xiang Mian, Ning Qian

Sichuan University, No.24 South Section 1, Yihuan Road, Chengdu, 610065, Sichuan, China · Hubei Minzu University, 39 Xueyuan Road, Enshi, 445000, Hubei, China

Source-Free Domain AdaptationVLM Adaptation
AskPDF
Jul 20, 2026

LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation

Source-free universal domain adaptation (SF-UniDA) adapts a pre-trained source model to an unlabeled target domain under both covariate and label shifts, without access to source data. However, existing SF-UniDA methods rely on inefficient techniques such as threshold tuning and clustering. Foundation models (FMs), known for their generalization and zero-shot capabilities, remain underexplored in SF-UniDA. In this paper, we propose a framework that leverages foundation models (LFM) for SF-UniDA. We use a vision-language model (VLM) to compute similarities between target samples and text labels, including those for unknown classes generated by prompting a large language model. The label shift type is determined by analyzing the coefficient of variation of a similarity-based sample-level score. Unknown samples are identified using a binary Gaussian mixture model fitted to another similarity-based metric. Under a consensus strategy, the pseudo-labels generated by the VLM are refined by the target model initialized with the pre-trained source model, integrating knowledge from both the source domain and foundation models. Finally, these refined pseudo-labels are used to train the target model. Extensive experiments across all possible label shifts and multiple benchmarks demonstrate the effectiveness and superiority of our proposed LFM framework. Our code is available at https://github.com/iamjingli/LFM.

Jing Li, Pan Liu, Meng Zhao +7

School of Computer Science and Engineering, Tianjin University of Technology, Tianjin 300384, China · Engineering Research Center of Learning-Based Intelligent System, Ministry of Education of the People’s Republic of China, Tianjin University of Technology, Tianjin 300384, China · School of Artificial Intelligence, Tianjin University, Tianjin 300350, China +1

Source-Free Domain AdaptationUnsupervised Domain AdaptationDomain Adaptation
AskPDF
Aug 5, 2026

COSMO: Consensus-Driven Shift Modulation for Source-Free Domain Adaptation

Source-free domain adaptation (SFDA) adapts a source-trained model to an unlabeled target domain without source data, a practical setting under privacy or storage constraints. Yet its self-generated supervision can reinforce source bias under substantial domain shifts. Pretrained vision-language models (VLMs) offer complementary semantic knowledge, but the relative reliability of the source model and VLM varies across target samples. Existing cross-model guidance does not explicitly account for this variation and may overwrite valid source-derived evidence under conflict, a failure we term source-derived evidence forgetting. We formulate VLM-guided SFDA as a sample-wise reliability-allocation problem and propose Consensus-Driven Shift Modulation (COSMO). COSMO replaces expert-to-expert guidance with co-adaptation through an anchored shared consensus. It first forms a sample-specific initial consensus that favors the more concentrated prediction. During adaptation, COSMO re-aggregates both branches' evolving evidence and regulates how far the resulting consensus moves from its initial anchor based on consensus uncertainty and training progress. This keeps the shared supervision anchored yet adaptive. Across four benchmarks, COSMO achieves state-of-the-art performance under matched VLM backbones. Further analyses indicate that it better balances the retention of valid source-derived evidence with the absorption of complementary VLM evidence.

Bo Li, Junjie Peng, Xiaohua Xie +1

School of Computer Science and Engineering, Sun Yat-sen University

Source-Free Domain AdaptationVLM Adaptation
AskPDF