Machine unlearning removes the influence of specific training examples to comply with privacy regulations. Full retraining gives strong guarantees but is computationally prohibitive at scale; approximate methods are efficient but lack formal guarantees. SISA (Sharded, Isolated, Sliced, and Aggregated training) achieves certified deletion through data sharding and independent retraining, but random partitioning scatters forget requests across most shards, and isolated retraining degrades utility. We introduce IDSISA (Influence-Distilled SISA), which uses influence functions to concentrate forget-prone samples into few shards and distillation from unaffected shards to recover utility lost to isolation. The deletion guarantee inherited from SISA's retraining step remains exact and requires no probabilistic qualification; the distillation-based utility recovery is a separately quantified, empirically bounded approximation (εcert≤0.015 in expectation). We further formalize certification under sequences of repeated deletion requests, where cross-shard distillation can create dependencies between shards; we prove that a closure-based cascading retraining procedure restores the deletion guarantee to the full aggregated model under any deletion sequence. Experiments on ResNet-18/CIFAR-10, ResNet-18/CIFAR-100, BiLSTM/IMDB, TCN/IMDB, and a fully connected network on Fashion-MNIST show 14−91% better selective forgetting, up to 50% better utility retention in relative terms (with consistent percentage-point gains on the other architectures, where a relative baseline is undefined), and unlearning-time changes ranging from −6.7% to +6.4% depending on architecture, relative to SISA.
Figures & tables
Method Class
Certified?
Selective Forgetting
Knowledge Pres.
Efficiency
Full Retraining
Yes
Strong
Strong
Low
Approx. Scrubbing
No
Moderate
Poor
High
Descent-to-Delete
No
Moderate
Poor
High
SISA (Random Shards)
Yes
Weak
Poor
High
Federated Unlearning
Yes
Moderate
Moderate
High
IDSISA (Proposed)
Yes †
Strong
Strong
High
TABLE I: Landscape of machine unlearning approaches. † Retraining component is deterministically certified; distillation introduces a bounded empirical approximation ( ϵcert≤0.015 , mean).
Symbol
Definition
D
Full training dataset {(xi,yi)}i=1N
Df
Forget set: samples to be unlearned
Dr
Retain set: Dr=D∖Df
Dm
Data assigned to shard m
θm
Model parameters for shard m
θm(k)
Checkpoint after slice k of shard m
TABLE II: Key notation used throughout this paper.
Operation
Theoretical
F-MNIST (ANN)
CIFAR-10 (CNN)
IMDB (LSTM)
Influence Calculation
O(N⋅p⋅L)
12s
38s
28s
PCA Reduction
O(N⋅d2)
1s
3s
2s
K-means Clustering
O(N⋅K⋅d′)
8s
15s
11s
Initial Training (SISA)
O(N⋅E⋅F)
8.9 min
59.9 min
120.16s
Initial Training (IDSISA)
8.9 min
66.1 min
138.32s
Unlearning (IDSISA)
O(∣Df∣⋅∣A∣⋅E′)
9.5 min
81.2 min
93s
TABLE III: Theoretical complexity and empirical runtime breakdown across architectures (mean over 5 seeds, 42–46). “Initial Training” is disaggregated per architecture, consistent with Table VI . † Speedup Factor (measured) is Full Retraining time divided by IDSISA Unlearning time for a single request; distinct from the theoretical M/∣A∣ speedup shown here and from the SISA-relative Unlearn Speedup column of Table IV .
Fig. 1: IDSISA’s three phases: influence-aware partitioning (I), slice-based shard training (II), and distillation-guided unlearning using unaffected shards as teachers (III).
Architecture
Dataset
Params
Forgetting Improv.
Preservation Improv. ‡
Unlearn Speedup
LSTM (BiLSTM)
IMDB Sentiment
4.2M
+91%
+3.75 pts
+2.1%
(3.94 vs 2.06%)
(+0.65 vs -3.10%)
(93 vs 95s)
CNN (ResNet-18)
CIFAR-10 Image
11.2M
+33.9%
+50%
+6.4%
(12.3 vs 18.6%)
(0.8 vs 1.6%)
(81.2 vs 86.7m)
ANN (Fully-Conn.)
Fashion-MNIST
203K
+14.3%
+0.03 pts
-6.7%
(-9.57 vs -8.37%)
(+0.03 vs 0.00%)
(534 vs 570s)
TABLE IV: Overall performance comparison: SISA vs. IDSISA across five architecture/dataset pairs (mean over 5 seeds, 42–46). ‡ Preservation Improv. is a relative % change where the baseline is non-zero and same-signed (CNN/CIFAR-10, CNN/CIFAR-100); where the baseline is negative, near-zero, or (CIFAR-100) not attributable to the unlearning mechanism, a relative % is undefined or not meaningful, so we report the raw percentage-point (pt) change instead. § Both methods report 0.0% forget-superclass accuracy on CIFAR-100; Section VI-F explains why this is not reported as a forgetting-quality ratio.
Method
Certified?
Forget Acc (%) ↓
Retain Acc (%) ↑
Test Acc (%) ↑
MIA on Df
Full Retrain (Oracle)
Yes
∼ 10.0
∼ 93.4
∼ 93.4
Calibrated
SalUn [ 21 ]
No †
∼ 0–6
∼ 88–92
∼ 88–92
Low
SCRUB [ 17 ]
No †
∼ 5–8
∼ 83–85
∼ 83–85
Moderate
SSD [ 18 ]
No †
∼ 2–8
∼ 82–86
∼ 82–86
Calibrated
SISA [ 1 ]
Yes
18.6
83.2
82.3
0.134
IDSISA
Yes ‡
12.3
84.7
83.1
0.132
TABLE V: Comparison with prior methods on CIFAR-10/ResNet-18, class-level forgetting ( ∼ 5,000 samples). † No formal deletion guarantee. ‡ See Table I : retraining certified, distillation gap ϵcert≤0.015 .
TABLE XIII: Impact of IDSISA components, CNN/CIFAR-10 (mean over 5 seeds, 42–46).
Architecture
ϵcert (mean)
Max Gap (99%ile)
Theoretical Bound (mean)
CNN (CIFAR-10)
0.011
0.024
0.015
LSTM (IMDB)
0.008
0.019
0.015
ANN (Fashion-MNIST)
0.013
0.027
0.015
Average
0.011
0.023
0.015
TABLE XIV: Certification gaps across architectures (mean over 5 seeds, 42–46). ϵcert measures distillation approximation only, not deletion correctness.
Fig. 2: Distribution of ϵcert , CNN/CIFAR-10, 10K test samples. The theoretical bound (dashed) reflects Theorem 1 ’s mean-case prediction, not a uniform bound over the tail.
TABLE XVI: Amortized cost vs. requests, CIFAR-10/CNN, single-deletion-per-request regime. Reset row applies a periodic full-shard refresh every K=5 requests (Remark 6, Appendix C).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Req.
∣A0∣
∣C∣
Cascade?
Scratch
Cum.
fallback?
edges
1
6
6
No
No
0
2
5
7
Yes
No
2
3
7
10
Yes
Yes
0
4
8
8
No
No
0
5
5
9
Yes
No
4
Appendix
TABLE XVII: Supplementary trace: 8 sequential class-structured requests, no resets (CIFAR-10/ResNet-18, M=10 ). Cascades on requests 2, 3, 5, 6, 8; full collapse to ∣C∣=M on requests 3 and 6.
Method
Certified Deletion
Selective Forgetting
Utility Pres.
Multi- Arch
NLP Support
MI Safe
Amor- tized
Influence Guided
Reference
Full Retraining
✓
Strong
Strong
✓
✓
✓
×
×
Oracle
Gradient Ascent [ 25 ]
×
Moderate
Poor
∼
∼
×
×
×
2022
Fine-Tuning [ 6 ]
×
Weak
Good
∼
∼
×
×
×
2020
Fisher Forgetting [ 6 ]
∼ (DP)
Moderate
Poor
×
×
×
×
∼
2020
NegGrad+ [ 17 ]
×
Moderate
Moderate
∼
∼
∼
×
×
2023
SCRUB [ 17 ]
×
Strong
Moderate
∼
×
∼
×
×
2023
Appendix
TABLE XVIII: Cross-dataset qualitative comparison. “Multi-Arch” = validated across CNN, RNN, and ANN. “MI Safe” = MI susceptibility calibrated to full-retrain level. “Amortized” = efficiency improves with repeated requests. ✓ / × / ∼ = satisfied / not / partially.
Method
Type
Certified?
Forget Acc (%) ↓
Retain Acc (%) ↑
Test Acc (%) ↑
MIA Conf. on Df↓
Full Retraining
Exact
Yes
∼ 10.0
∼ 93.4
∼ 93.4
∼ 0.68
Gradient Ascent [ 25 ]
Weight mod.
No †
∼ 0–6 ∗
∼ 74–80 ∗
∼ 63–75 ∗
Low
Fine-Tuning [ 6 ]
Weight mod.
No †
∼ 60–80 ∗
∼ 83–85 ∗
∼ 82–84 ∗
High
NegGrad+ [ 17 ]
Gradient
No †
∼ 5–10 ∗
∼ 82–84 ∗
∼ 82–84 ∗
Moderate
Fisher Forgetting [ 6 ]
DP noise
Approx. DP
∼ 3–8 ∗
∼ 78–82 ∗
∼ 78–82 ∗
Low
SSD [ 18 ]
FIM dampening
No †
∼ 2–8 ∗
∼ 82–86 ∗
∼ 82–86 ∗
Calibrated
Appendix
TABLE XIX: Quantitative comparison, CIFAR-10/ResNet-18, class-level forgetting ( ∼ 5,000 samples). † No formal deletion guarantee. ‡ Retraining certified; distillation gap ϵcert≤0.015 (Table I, main paper). ∗ Representative figures from published benchmarks [ 17 , 21 , 18 , 6 ] ; exact figures vary by implementation.
Model
Df Conf.
Df TPR
Dtrain TPR
Threshold
Val Acc
SISA (Unlearned)
0.134
≈ 0.42
≈ 0.60
≈ 0.139
0.62
IDSISA (Unlearned)
0.132
≈ 0.39
≈ 0.59
≈ 0.139
0.62
Full Retrain ( Dr )
0.681
0.83
N/A
0.738
N/A
Appendix
TABLE XX: Full membership-inference results (CIFAR-10/ResNet-18, shadow model + threshold attack).
Machine unlearning removes the influence of selected data or concepts from trained models while preserving useful behavior and avoiding full retraining.