Continual model updating has become a common practice in the deployment of machine learning systems. However, the dynamic training behaviors give rise to new risks not sufficiently addressed by existing studies. In this paper, we present the first systematic study of temporal privacy risk in continually updated deep learning models. We formalize the notion of temporal membership in continually updated models. We then define the temporal membership inference attack, which seeks to infer whether and when individual samples were involved in model training, at both timestamp-level and interval-level granularities. To this end, we propose TMIA, a novel black-box attack that leverages bidirectional LSTMs to infer temporal membership from model outputs observed across successive updates. Through extensive experiments on three language models and three image classification models, we show that TMIA is highly effective and significantly outperforms existing baselines. Finally, we evaluate three potential defenses, revealing the trade-off between privacy protection and model utility.
Figures & tables
Study
Operation
Frequency
Attack Input
( Jagielski et al., 2023 )
INS only
One-time
fi,fi+1
( Naderloui et al., 2025 ; Hayes et al., 2024 ; Chen et al., 2021 )
DEL only
One-time
fi,fi+1
( Michel et al., 2026 )
INS only
One-time
f0,f1,…,fN
Ours
INS & DEL
Repeatedly
f0,f1,…,fN
Table 1. Comparison between TMIA and existing MIAs against models with updates. INS: Insertion; DEL: deletion; fi,fi+1 : consecutive model snapshot pair; f0,f1,…,fN : Full sequence of model snapshots.
Figure 1. An example of the ground-truth and inferred temporal membership intervals.
Figure 2. Entry and exit points predicted by three straw-man approaches (Pythia-70M model, SST-5 datasets). The target sample was initially present in the training set. It was removed at the 36th timestamp, and re-inserted at the 70th timestamp.
Figure 3. Illustration of the sequence labeling model.
Figure 4. An example of membership interval construction.
Model
Dataset
TMIA (Ours)
Timestamp-wise MIA
Update-wise MIA
T-Pre
T-Rec
T-F1
T-Pre
T-Rec
T-F1
T-Pre
T-Rec
T-F1
Pythia-70M
SST-5
88.78
94.06
91.34
69.26
88.21
77.59
52.01
65.25
57.88
News20
82.55
84.58
83.55
57.33
79.44
66.60
29.01
59.81
39.07
MNLI
83.67
91.19
87.27
72.91
94.16
82.18
45.35
75.41
56.64
T5
SST-5
84.73
92.19
88.30
61.53
90.94
73.40
46.35
89.87
61.16
News20
80.51
88.10
84.13
57.25
81.31
67.19
31.78
63.26
42.30
Table 2. Timestamp-level accuracy of TMIA and baselines. The best T-F1 results for each setting are marked in gray .
Model
Dataset
TMIA (Ours)
Timestamp-wise MIA
Update-wise MIA
I-Pre
I-Rec
I-F1
I-Pre
I-Rec
I-F1
I-Pre
I-Rec
I-F1
Pythia-70M
SST-5
80.89
87.70
84.16
56.47
57.57
57.01
46.93
68.35
55.65
News20
75.04
79.58
77.24
47.52
50.85
49.13
30.66
64.35
41.53
MNLI
78.84
76.63
77.72
46.50
59.73
52.29
39.16
74.89
51.43
T5
SST-5
68.70
77.97
73.04
53.23
61.25
56.96
29.85
87.63
44.53
News20
74.15
71.69
72.90
50.31
51.92
51.10
35.70
68.67
46.98
Table 3. Interval-level accuracy of TMIA and baselines. The best I-F1 results for each setting are marked in gray .
Figure 5. Timestamp-level performance of TMIA when the shadow model mismatches the target model. The horizontal and vertical axes correspond to the target and shadow configurations, respectively, including the model, dataset and update method.
Figure 6. Performance of TMIA under various parameter configurations (Pythia-70M, SST-5)
Model
Dataset
TMIA (Ours)
Timestamp-wise MIA
Update-wise MIA
T-Pre
T-Rec
T-F1
T-Pre
T-Rec
T-F1
T-Pre
T-Rec
T-F1
Llama-3B
Squad
99.48
81.59
89.65
53.87
95.70
68.79
32.87
47.11
38.72
Simpleqa
98.25
92.15
95.25
59.05
99.98
74.06
57.83
47.56
52.19
Llama-8B
Squad
96.22
85.16
90.35
50.61
94.98
65.88
32.34
46.66
38.21
Simpleqa
95.61
75.45
84.34
55.08
99.95
70.88
56.73
45.66
50.6
Table 4. Timestamp-level accuracy of TMIA and baselines for language models. The best results for each setting are marked in gray .
Timestamp-level
Interval-level
Updates
T-Pre
T-Rec
T-F1
I-Pre
I-Rec
I-F1
Uncorrelated
88.81
91.51
90.14
72.10
89.92
80.03
Correlated insertion
88.50
91.70
90.07
71.50
90.60
79.92
Correlated removal
88.30
91.20
89.73
71.40
89.80
79.55
Correlated (both)
87.60
91.80
89.65
70.20
91.30
79.37
Table 5. Performance of TMIA under correlated model updates (SST-5, Pythia-70M).
Timestamp-level
Interval-level
Setting
T-Pre
T-Rec
T-F1
I-Pre
I-Rec
I-F1
Full precision
88.81
91.51
90.14
72.10
89.92
80.03
8-bit
83.64
77.26
80.32
70.61
74.53
72.52
4-bit
75.72
54.33
63.27
62.52
62.49
62.50
Table 6. Performance of TMIA against reduced-precision model outputs by quantization (Pythia-70M model).
Access to
Timestamp-level
Interval-level
model updates
T-Pre
T-Rec
T-F1
I-Pre
I-Rec
I-F1
Full sequence
88.81
91.51
90.14
72.10
89.92
80.03
50% of snapshots
69.07
76.98
72.82
60.54
69.39
64.66
25% of snapshots
40.42
60.55
48.48
37.96
35.10
36.47
Table 7. Performance of TMIA under partial access to model updates (Pythia-70M model).
Dataset/Model
Defense
Timestamp-level
Interval-level
Model Utility
T-Pre
T-Rec
T-F1
I-Pre
I-Rec
I-F1
Train. Acc
Test. Acc
SST-5/Pythia-70M
W/o defense
88.78
94.06
91.34
80.89
87.70
84.16
99.18
43.26
Label-only output
84.79
91.44
87.99
67.90
83.39
74.85
99.18
43.25
Dropout p=30%
80.71
71.77
75.36
73.63
72.84
73.22
65.12
40.13
Dropout p=50%
76.39
46.99
58.18
64.15
43.51
51.82
60.11
41.24
DP ϵtotal=30
71.60
24.03
35.99
27.47
14.93
19.35
54.67
40.11
Table 8. Performance of three defense mechanisms against TMIA .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Modality
Feature dim.
# labels
# train samples
# test samples
SST-5
Text
512
5
8,544
2,210
News20
Text
512
20
11,314
7,532
MNLI
Text
512
3
50,000
9,815
CIFAR-10
Image
3072
10
50,000
10,000
SVHN
Image
3072
10
73,257
26,032
CINIC-10
Image
3072
10
90,000
90,000
Appendix
Table 9. Statistics of datasets
Model
# Layers
Layer width
# of Parameters
Pythia-70M
6
512
70M
GPT-2 Small
12
768
124M
T5
12
512
220M
ResNet-18
18
{64,128,256,512}
11.7M
SimpleCNN
4
{32,64,128,128}
1.2M
MobileNet
28
{32–1024}
4.2M
Appendix
Table 10. Architectural details of models used in the experiments
Figure 7. Change of confidence score across model snapshots (SimpleCNN model, CIFAR-10 dataset). The target sample was initially present in the training set. It was removed at the 27th timestamp, and re-inserted at the 48th timestamp.
Figure 8. Accuracy of inferring entry and exit points
Figure 9. Interval-level attack Performance of TMIA when the shadow model mismatches the target model. The horizontal and vertical axes correspond to the target and shadow configurations, respectively, including the model, dataset, update method, and update volume.
Figure 10. Performance of TMIA under various parameter configurations (Pythia-70M, News20)
Model
Dataset
TMIA (Ours)
Timestamp-wise MIA
Update-wise MIA
I-Pre
I-Rec
I-F1
I-Pre
I-Rec
I-F1
I-Pre
I-Rec
I-F1
Llama-3B
Squad
97.37
82.59
89.37
38.42
84.67
52.86
34.30
10.90
16.54
Simpleqa
97.23
92.89
95.01
34.98
59.15
43.96
60.01
8.84
15.41
Llama-8B
Squad
94.63
87.14
90.73
35.32
82.68
49.50
33.86
10.55
16.09
Simpleqa
96.46
73.22
83.25
31.71
49.91
38.78
58.82
8.35
14.62
Appendix
Table 11. Interval-level acc of TMIA and baselines on generative models. The best results for each setting are marked in gray .
Deep learning models for time series imputation are now essential in fields such as healthcare, the Internet of Things (IoT), and finance. However, their deployment raises critical privacy concerns. Beyond the well-known issue of unintended memorization, which has been extensively studied in generative models, we demonstrate that time series models are vulnerable to inference attacks in a black-box setting. In this work, we introduce a two-stage attack framework comprising: (1) a novel membership inference attack based on a reference model that improves detection accuracy, even for models robust to overfitting-based attacks, and (2) the first attribute inference attack that predicts sensitive characteristics of the training data for timeseries imputation model. We evaluate these attacks on attention-based and autoencoder architectures in two scenarios: models that are trained from scratch, and fine-tuned models where the adversary has access to the initial weights. Our experimental results demonstrate that the proposed membership attack retrieves a significant portion of the training data with a tpr@top25% score significantly higher than a naive attack baseline. We show that our membership attack also provides a good insight of whether attribute inference will work (with a precision of 90% instead of 78% in the genral case).
Faiz Taleb, Ivan Gazeau, Maryline Laurent
EDF SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris, 91120 Palaiseau, France · EDF, 91120 Palaiseau, France · SAMOVAR, Télécom SudParis, Institut Polytechnique de Paris
Modern AI models are not static. They go through multiple updates in their lifecycles. We propose to design Sequential Membership Inference (SeMI) attacks leading to tighter privacy audits by exploiting the sequence of models and injecting a target canary at a controlled insertion time. First, for empirical mean computation, we develop SeMI*, an optimal SeMI attack to identify the presence of a target inserted at a specific insertion step. We derive the power of SeMI* to show that accessing the model sequence yields more powerful MI attacks than scrutinising only the final model. SeMI* exhibits an isolation property. Its power depends on the statistics obtained right before and after insertion of the target. Leveraging this insight, we develop practical white-box (accessing model gradients) and black-box (accessing loss) SeMI attacks against models trained with (DP-)SGD. Across datasets and models trained with (DP-)SGD, our experiments show that SeMI attacks achieve higher power than snapshot-independent baselines, and yield tighter privacy audits thanks to (a) control over the insertion time and (b) observations across the model sequence. Code is available at https://github.com/Thomick/pysemi.
Thomas Michel, Debabrota Basu, Emilie Kaufmann
Univ. Lille, Inria, CNRS, Centrale Lille UMR 9189-CRIStAL, France
The pervasive deployment of deep learning models across critical domains has concurrently intensified privacy concerns due to their inherent propensity for data memorization. While Membership Inference Attacks (MIAs) serve as the gold standard for auditing these privacy vulnerabilities, conventional MIA paradigms are increasingly constrained by the prohibitive computational costs of shadow model training and a precipitous performance degradation under low False Positive Rate constraints. To overcome these challenges, we introduce a novel perspective by leveraging the principles of model reprogramming as an active signal amplifier for privacy leakage. Building upon this insight, we present ReproMIA, a unified and efficient proactive framework for membership inference. We rigorously substantiate, both theoretically and empirically, how our methodology proactively induces and magnifies latent privacy footprints embedded within the model's representations. We provide specialized instantiations of ReproMIA across diverse architectural paradigms, including LLMs, Diffusion Models, and Classification Models. Comprehensive experimental evaluations across more than ten benchmarks and a variety of model architectures demonstrate that ReproMIA consistently and substantially outperforms existing state-of-the-art baselines, achieving a transformative leap in performance specifically within low-FPR regimes, such as an average of 5.25% AUC and 10.68% TPR@1%FPR increase over the runner-up for LLMs, as well as 3.70% and 12.40% respectively for Diffusion Models.