Pith. sign in

REVIEW 5 major objections 6 minor 27 references

POLARIS: Explainable Artificial Intelligence for Mitigating Power Side-Channel Leakage

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that an explainable-AI masking flow can cut power side-channel leakage more than the state-of-the-art full-masking baseline while masking only half the leaky gates and running about six times faster.

desk verdict A plausible ML-guided masking flow whose headline result is in-sample because the model family is chosen on the same benchmarks used for the final comparison. read the letter →

arxiv 2507.22177 v1 pith:TCTL7STN submitted 2025-07-29 cs.CR

Tanzim Mahfuz , Sudipta Paria , Tasneem Suha , Swarup Bhunia , Prabuddha Chakraborty This is my paper · ORCID
classification cs.CR
keywords powerside-channelattackmaskingexplainableAISHAPleakageassessmenthardwaresecuritygate-levelnetlistdesign-for-security
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Power side-channel attacks infer secret data from a chip's dynamic power trace, and inserting masking gates is a common defense, but choosing which gates to mask is slow when each candidate requires leakage simulation. POLARIS proposes replacing those per-gate simulations with a machine-learning model trained on the local structure of a gate-level netlist. The model is trained on smaller open-source circuits using an unsupervised labeling scheme that randomly inserts masks and records whether measured leakage improved; it is then applied to larger, unseen designs. The paper reports that this flow outperforms the state-of-the-art masking baseline in leakage reduction, runtime, and area/power/delay overhead, and that the explanation step yields human-readable masking rules. The headline result is that masking only 50% of the leaky gates already gives 45.72% leakage reduction, slightly more than the baseline's 45.46% at full masking.

What carries the argument

The load-bearing object is a structural fingerprint of each gate: the gate's type one-hot encoded together with the types and adjacency of gates within a radius-L neighborhood, collected by breadth-first search over the netlist graph. Training labels come from random mask insertions whose TVLA leakage reduction is compared against a threshold, and the trained model is an AdaBoost classifier, chosen after comparing random forests and XGBoost. SHAP, a game-theoretic method that assigns each feature a Shapley value, attributes predictions to structural features and produces the human-readable rules. The masking step sorts all gates by predicted benefit, takes the top-ranked gates up to the mask budget, and replaces them with composite masking gates, then verifies leakage with TVLA.

What would settle it

Train the model on the smaller circuits, hold out one of the large designs entirely without using it for model selection, and compare POLARIS's leakage reduction on that held-out design against random masking with the same number of gates; if random masking matches or beats the model, the learned structural signal is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that power side-channel mitigation can be driven by a learned model rather than by exhaustive leakage testing. For each gate, POLARIS builds a structural fingerprint that encodes the gate's type and the types and connectivity of neighboring gates within a fixed radius, and labels random mask insertions as good or bad according to whether TVLA-measured leakage drops by at least a threshold. An AdaBoost model trained on these labels ranks gates by predicted masking benefit, and the top-ranked gates are replaced with composite masking gates. On the reported benchmark suite, the resulting masked designs show 45.72% leakage reduction at 50% masking, 50.88% at 75%, and 54.09% at full masking, compared with 45.46% for the full-masking baseline, with roughly six times lower runtime and about one-third lower area, power, and delay overhead. The paper also claims that SHAP-based rule extraction turns the model into explicit, human-readable 'mask' and 'do not mask' rules.

Load-bearing premise

The framework's advantage rests on the assumption that the mapping from a gate's local neighborhood structure to masking benefit transfers from small training circuits to larger, unseen designs, and the reported numbers are not checked on a held-out validation split.

Editorial extensions

If this is right

  • Masking decisions no longer require a TVLA simulation per candidate gate, so the mitigation step becomes a single model-inference pass over the netlist; reported runtimes drop from hundreds of seconds to tens of seconds on the test designs.
  • Because 50% masking already meets or exceeds the baseline's full-masking leakage reduction, the same protection can be delivered at materially lower area, power, and delay cost.
  • The extracted SHAP rules give a designer an auditable list of when to mask and when not to mask, making the mitigation strategy explainable rather than a black-box placement.
  • Since training data are generated synthetically and labeling is automatic, the same flow can be retrained for other gate libraries, other masking styles, or other design families without collecting an external dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper bypasses TVLA only at inference time: TVLA still generates the training labels, so a fully TVLA-free flow would need a leakage proxy for labeling as well.
  • If the structural-transfer assumption holds, the same model class could be applied earlier in the design flow, for instance to prune masking candidates at synthesis time, or retrained for FPGA lookup-table netlists as the paper suggests.
  • A directly testable extension is to hold out one of the large evaluation designs during model selection and report its leakage reduction; that would quantify how much of the reported average comes from genuine transfer rather than from tuning on the same eleven designs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript presents POLARIS, a design-for-security flow that trains an ML classifier on structural features of gates (gate type, local connectivity, and a radius-L neighborhood), explains the model with SHAP, and uses the model to select gates for replacement by masked logic gates. The authors evaluate POLARIS on 11 EPFL/MIT-CEP benchmarks and report that POLARIS with 50% masking achieves 45.72% average leakage reduction versus 45.46% for VALIANT with full masking, with lower area/power/delay overhead and roughly 6x shorter runtime. The paper claims that POLARIS is faster, more effective, and lower-overhead than the state-of-the-art VALIANT flow.

Significance. If the evaluation were out-of-sample and reproducible, POLARIS would be a useful contribution: it addresses a real scalability bottleneck in TVLA-based masking flows, provides a transparent rule-extraction mechanism via SHAP, and is evaluated on diverse open-source benchmarks. The paper's strengths include an automated training-data generation scheme that avoids hand-labeled datasets, integration of the flow into an ASIC design process, and the use of standard benchmark suites. However, the current evaluation does not yet establish the headline claims, because model selection is performed on the same benchmarks used for the final reported numbers and because no variance information is reported.

major comments (5)
  1. [V.B, Table III] Table III selects AdaBoost as the final model because its 54.09% average leakage reduction is the highest among Random Forest, XGBoost, and AdaBoost, and that average is computed over exactly the same 11 EPFL/MIT-CEP benchmarks on which the headline results in Table II are reported. No held-out validation set is used before model selection, and Section V.A also justifies hyperparameters (L=7, θr=0.70) with reference to outcomes on these benchmarks. Consequently, the comparison against VALIANT in Table II is in-sample with respect to model selection and hyperparameter choice, and the transfer-learning claim from ISCAS-85 to larger designs is untested. Please add a separate validation set or nested cross-validation, and report results on designs that were not used for any model or hyperparameter selection.
  2. [IV.A, Algorithm 1] Algorithm 1's labeling procedure (lines 2, 9, and 12) computes a leakage estimate before and after random masking and labels a gate 'good' if rRatio >= θr, where the leakage values are TVLA values as described in Section II.A and Section V.A. This directly contradicts the statement in the Abstract and Introduction that POLARIS 'does not rely on TVLA for leakage estimation and mitigation through masking.' At most, TVLA is bypassed at inference/masking time (Algorithm 2), not in the training-data generation or evaluation stages. The claim should be scoped accordingly, and the cost of TVLA labeling on training designs should be included in the scalability analysis.
  3. [V.A, Table II] The reported 6x speedup of POLARIS over VALIANT appears to compare per-benchmark masking time only, while Section V.A states that model training takes approximately 40 minutes and the training-data generation itself invokes TVLA on the ISCAS-85 designs. Table II's 'POLARIS' time column seems to exclude this one-time training and labeling cost. To support an end-to-end speedup claim, please report training time, labeling time, and per-design inference/masking time separately, and either include the one-time cost in the comparison or explicitly state that the speedup is for the masking phase only.
  4. [V, Tables II and III] All leakage-reduction numbers are single runs with no error bars, confidence intervals, or indication of run-to-run variability. This matters particularly for the headline comparison: the claimed advantage of POLARIS at 50% masking (45.72%) over VALIANT at full masking (45.46%) is only 0.26 percentage points, and Algorithm 1's random gate-selection process introduces stochasticity into training-data generation. Please repeat the experiments multiple times with different random seeds and report mean and standard deviation (or an equivalent measure of dispersion).
  5. [V.C, Table II] The manuscript does not state how the VALIANT baseline was obtained—whether the numbers were re-run under the same synthesis, trace count, and simulation setup as POLARIS, or taken from the VALIANT paper. Since the central comparison is against VALIANT, please specify the baseline setup (tool versions, technology library, trace count, TVLA configuration, and whether all 11 benchmarks were processed with the same flow) so that the comparison is meaningful.
minor comments (6)
  1. [Throughout] The text and tables repeatedly use 'V ALIANT' with an extra space; this should be 'VALIANT'. Also, the area unit in Table IV should be µm², not 'µM²'.
  2. [Table V] Rule B in Table V contains the stray word 'blue' in the condition column; this appears to be a leftover and should be removed.
  3. [Algorithm 2] Algorithm 2, line 8 sorts C in descending order, but C contains tuples (Ppred, i); the tie-breaking rule for the second element should be specified to make the procedure deterministic.
  4. [IV.A] Section IV.A describes the data-generation process as 'unsupervised,' but Algorithm 1 assigns binary labels using a leakage-reduction threshold; this is supervised labeling, and the terminology should be clarified.
  5. [Table I] Table I claims adversarial-attack immunity ('Adversarial Attacks: No') for POLARIS, but no adversarial robustness experiments are reported; this claim should be softened or removed.
  6. [IV.A, Fig. 2] The structural feature vector is defined only at a high level (adjacency matrix, one-hot encoding, BFS with radius L); a precise, formal definition of the feature vector and the encoding would be needed for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline leakage-reduction numbers are in-sample with respect to model selection: AdaBoost is chosen because it maximizes the average leakage reduction over the same 11 benchmarks later reported as POLARIS's result.

  1. fitted input called prediction [Section V.B, Table III, and Table II]
    "AdaBoost outperforms the other models with a 54.09% leakage reduction on average, making it the chosen model for the remaining experiments."

    The ML model family is selected by comparing average leakage reduction over the same 11 EPFL/MIT-CEP benchmarks that later form the evaluation set for the headline claims. The number used to choose AdaBoost (54.09% average) is exactly the number reported as POLARIS's full-mask average in Table II. Thus the central 'prediction' that POLARIS with full masking achieves 54.09% leakage reduction is the model-selection statistic itself, not an independent out-of-sample result. The training netlists are separate ISCAS-85 designs, so individual gate predictions are not trained on these 11 netlists, but the algorithm choice is fitted to the reported outcome.

full rationale

The main circular step is the test-set model selection: Table III compares Random Forest, XGBoost, and AdaBoost on the exact 11 benchmarks later reported in Table II, and AdaBoost is chosen because its 54.09% average leakage reduction is the highest. That same 54.09% is then presented as POLARIS's full-mask result. This is a fitted choice renamed as a prediction, because the reported outcome was used to select the model. I did not count the use of TVLA for training labels or final evaluation as circular per se: supervised learning with the same metric as the label and the evaluation is legitimate when the evaluation designs are held out. Here, however, the evaluation designs were not held out from model selection. I also did not treat the VALIANT comparison as circular solely because of author overlap, since VALIANT is an external prior work and its overlap is not load-bearing for the reduction. The more severe issue remains that the headline numbers are partly fitted to the benchmarks they claim to predict.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

All load-bearing assumptions are domain assumptions about the validity of TVLA, mask-gate effectiveness, and transferability of learned structural features. No invented physical entities are introduced. The main free parameters are hyperparameters chosen by hand or by test-set performance.

free parameters (7)
  • Theta_r (labeling threshold) = 0.70
    Gates reducing leakage by at least 70% are labeled 'good' for masking. Chosen by hand to avoid data imbalance.
  • L (neighborhood radius) = 7
    Number of neighboring gates considered in structural features. Set to 7 without sensitivity analysis.
  • Msize (mask size for training) = 200
    Number of randomly inserted masks per iteration during synthetic data generation. No sensitivity study reported.
  • itr (iterations) = 100
    Number of random masking iterations for data generation. Termination condition not validated.
  • Learning rate alpha = 0.01
    Used for XGBoost and AdaBoost. No sensitivity analysis given.
  • Model class = AdaBoost
    Selected because it achieves the highest average leakage reduction on the same 11 evaluation benchmarks (Table III), which is test-set model selection.
  • SMOTE / class weighting = SMOTE for Random Forest, weighted training for XGBoost and AdaBoost
    Handles label imbalance caused by Theta_r. A modeling choice with no ablation study.
assumptions (5)
  • domain assumption TVLA with |t| > 4.5 is a valid indicator of exploitable power side-channel leakage.
    Used to label training data and to measure all reported leakage reductions. If this metric is unreliable, the entire evaluation collapses.
  • domain assumption Masking gates (e.g., masked AND/OR from [4]) actually reduce leakage when inserted at selected gates.
    The framework assumes replacing a gate with a masked version reduces TVLA leakage. Standard in the field but not re-verified here.
  • domain assumption A gate's leakage effect is predictable from its type and local connectivity within radius L.
    Core inductive bias of the ML model. No independent evidence is given beyond the reported benchmarks.
  • domain assumption Models trained on small ISCAS-85 designs transfer to larger EPFL and MIT-CEP designs.
    The evaluation claims transfer learning, but no separate validation set or domain-gap analysis is provided.
  • domain assumption Random insertion of masks with a threshold comparison gives a sufficiently balanced and informative training set.
    The synthetic data generation depends on this assumption. No analysis of coverage or bias in the generated samples is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of POLARIS: Explainable Artificial Intelligence for Mitigating Power Side-Channel Leakage." pith.science (2026). https://pith.science/paper/TCTL7STN

@misc{pith2026250722177,
  author       = {Pith},
  title        = {Pith review of: POLARIS: Explainable Artificial Intelligence for Mitigating Power Side-Channel Leakage},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TCTL7STN}},
  note         = {Machine review of arXiv:2507.22177}
}
read the original abstract

Microelectronic systems are widely used in many sensitive applications (e.g., manufacturing, energy, defense). These systems increasingly handle sensitive data (e.g., encryption key) and are vulnerable to diverse threats, such as, power side-channel attacks, which infer sensitive data through dynamic power profile. In this paper, we present a novel framework, POLARIS for mitigating power side channel leakage using an Explainable Artificial Intelligence (XAI) guided masking approach. POLARIS uses an unsupervised process to automatically build a tailored training dataset and utilize it to train a masking model.The POLARIS framework outperforms state-of-the-art mitigation solutions (e.g., VALIANT) in terms of leakage reduction, execution time, and overhead across large designs.

Figures

Figures reproduced from arXiv: 2507.22177 by the authors.

Figure 1
Figure 1. AND and OR operation on masked data. C. Related Work Existing power side-channel leakage evaluation/mitigation approaches like CASCADE [7], Karna [8], VALIANT [6] in￾corporate TVLA analysis and hence suffers from high runtime and scalability issues. AI-based methods like DL-LA [9] and Netlist Whisperer [10] are often limited by the lack of training data and are vulnerable to adversarial attacks due to reliance on ex… view at source ↗
Figure 2
Figure 2. Overview of the POLARIS framework. power leakage. TVLA has major scalability concerns particu￾larly for large industry-grade designs that in turn hampers the scalability of frameworks such as Karna and VALIANT. We hypothesize, AI techniques might be able to predict the leakage (directly or in an integrated way) associated with individual gates faster and with a linear time complexity (with respect to design size). C… view at source ↗
Figure 3
Figure 3. SHAP waterfall plots generated by POLARIS Adaboost (ADB) model for power side channel defense. C. Comparison with State-of-the-Art Table II compares the performance of VALIANT framework [6] and POLARIS in terms of leakage per gate, total leakage reduction, and execution time. POLARIS is tested with 50%, 75%, and full (100%) masking sizes. Even with only 50% masking, POLARIS achieves a 45.72% leakage reduction, surpa… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: TVLA values before and after masking in des3 design. Gates exceeding threshold (±4.5) are considered as leaky. and full masking, POLARIS reduces leakage by 50.88% and 54.09%, respectively, outperforming VALIANT. Additionally, POLARIS operates 6x faster than VALIANT, de…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    Differential power analysis,

    P. Kocher, J. Jaffe, and B. Jun, “Differential power analysis,” inAdvances in Cryptology — CRYPTO’ 99 . Berlin, Heidelberg: Springer Berlin Heidelberg, 1999, pp. 388–397

    work page 1999

  2. [2]

    Side-Channel Leakage through Static Power – Should We Care about in Practice?

    A. Moradi, “Side-Channel Leakage through Static Power – Should We Care about in Practice?” Cryptology ePrint Archive, Paper 2014/025,

    work page 2014

  3. [3]

    A testing methodology for side channel resistance validation,

    G. Goodwill, B. Jun, J. Jaffe, and P. Rohatgi, “A testing methodology for side channel resistance validation,” in NIST Non-Invasive Attack Testing Workshop, vol. 7, 2011, pp. 115–136

    work page 2011

  4. [4]

    Combinational Logic Design for AES SubByte Transformation on Masked Data,

    E. Trichina, “Combinational Logic Design for AES SubByte Transformation on Masked Data,” Cryptology ePrint Archive, Paper 2003/236, 2003. [Online]. Available: https://eprint.iacr.org/2003/236

    work page 2003

  5. [5]

    Domain-Oriented Masking: Compact Masked Hardware Implementations with Arbitrary Protection Order,

    H. Gross, S. Mangard, and T. Korak, “Domain-Oriented Masking: Compact Masked Hardware Implementations with Arbitrary Protection Order,” in Proceedings of the 2016 ACM Workshop on Theory of Implementation Security, ser. TIS ’16, New York, NY , USA, 2016, p. 3

    work page 2016

  6. [6]

    V ALIANT: An EDA Flow for Side-Channel Leakage Evaluation and Tailored Protection,

    R. Sadhukhan, S. Saha, S. Paria, S. Bhunia, and D. Mukhopadhyay, “V ALIANT: An EDA Flow for Side-Channel Leakage Evaluation and Tailored Protection,” IEEE Transactions on Computers , vol. 73, no. 2, pp. 436–450, 2024

    work page 2024

  7. [7]

    Towards efficient and automated side-channel evaluations at design time,

    D. ˇSijaˇci´c, J. Balasch, B. Yang et al., “Towards efficient and automated side-channel evaluations at design time,” J Cryptogr Eng , vol. 10, pp. 305–319, 2020

    work page 2020

  8. [8]

    Karna: A Gate- Sizing based Security Aware EDA Flow for Improved Power Side- Channel Attack Protection,

    P. Slpsk, P. K. Vairam, C. Rebeiro, and V . Kamakoti, “Karna: A Gate- Sizing based Security Aware EDA Flow for Improved Power Side- Channel Attack Protection,” in 2019 IEEE/ACM International Confer- ence on Computer-Aided Design (ICCAD) , 2019, pp. 1–8

    work page 2019

Show all 27 references
  1. [9]

    DL-LA: Deep learning leakage assessment: A modern roadmap for SCA evaluations,

    T. Moos, F. Wegener, and A. Moradi, “DL-LA: Deep learning leakage assessment: A modern roadmap for SCA evaluations,” Cryptology ePrint Archive, Paper 2019/505, 2019. [Online]. Available: https://eprint.iacr.org/2019/505

    2019

  2. [10]

    Netlist whisperer: Ai and nlp fight circuit leakage!

    M. Nair, R. Sadhukhan, H. Pearce, D. Mukhopadhyay, and R. Karri, “Netlist whisperer: Ai and nlp fight circuit leakage!” in Proceedings of the 2023 Workshop on Attacks and Solutions in Hardware Security, ser. ASHES ’23. New York, NY , USA: Association for Computing Machinery, 2...

    2023

  3. [11]

    Leakage Assessment Methodology - a clear roadmap for side-channel evaluations,

    T. Schneider and A. Moradi, “Leakage Assessment Methodology - a clear roadmap for side-channel evaluations,” Cryptology ePrint Archive, Paper 2015/207, 2015. [Online]. Available: https://eprint.iacr.org/2015/ 207

    2015

  4. [12]

    Detecting Flawed Masking Schemes with Leakage De- tection Tests,

    O. Reparaz, “Detecting Flawed Masking Schemes with Leakage De- tection Tests,” in Revised Selected Papers of the 23rd International Conference on Fast Software Encryption - Volume 9783 , ser. FSE 2016. Berlin, Heidelberg: Springer-Verlag, 2016, p. 204–222

    2016

  5. [13]

    Masking against side-channel attacks: A formal security proof,

    E. Prouff and M. Rivain, “Masking against side-channel attacks: A formal security proof,” in Advances in Cryptology – EUROCRYPT 2013. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 142–159

    2013

  6. [14]

    Is poisoning a real threat to llm alignment? maybe more so than you think,

    P. Pathmanathan, S. Chakraborty, X. Liu, Y . Liang, and F. Huang, “Is poisoning a real threat to llm alignment? maybe more so than you think,” arXiv preprint arXiv:2406.12091 , 2024

    2024 arXiv

  7. [15]

    Data poisoning attacks against federated learning systems,

    V . Tolpegin, S. Truex, M. E. Gursoy, and L. Liu, “Data poisoning attacks against federated learning systems,” in Computer security–ESORICs 2020: 25th European symposium on research in computer security, ESORICs 2020, guildford, UK, September 14–18, 2020, proceedings, part i 2...

    2020

  8. [16]

    An overview of backdoor attacks against deep neural networks and possible defences,

    W. Guo, B. Tondi, and M. Barni, “An overview of backdoor attacks against deep neural networks and possible defences,”IEEE Open Journal of Signal Processing , vol. 3, pp. 261–287, 2022

    2022

  9. [17]

    Label-consistent backdoor at- tacks,

    A. Turner, D. Tsipras, and A. Madry, “Label-consistent backdoor at- tacks,” arXiv preprint arXiv:1912.02771 , 2019

    1912 arXiv

  10. [18]

    A unified approach to interpreting model predictions,

    S. Lundberg, “A unified approach to interpreting model predictions,” arXiv preprint arXiv:1705.07874 , 2017

    2017 arXiv

  11. [19]

    SAIL: Machine learning guided structural analysis attack on hardware obfuscation,

    P. Chakraborty, J. Cruz, and S. Bhunia, “SAIL: Machine learning guided structural analysis attack on hardware obfuscation,” in 2018 Asian Hardware Oriented Security and Trust Symposium (AsianHOST). IEEE, 2018, pp. 56–61

    2018

  12. [20]

    Learning Your Lock: Exploiting Structural Vulnerabilities in Logic Locking,

    P. Chakraborty, J. Cruz, R. Almawzan, T. Mahfuz, and S. Bhunia, “Learning Your Lock: Exploiting Structural Vulnerabilities in Logic Locking,” IEEE Design & Test , 2024

    2024

  13. [21]

    Why should i trust you? Explaining the predictions of any classifier,

    M. T. Ribeiro, S. Singh, and C. Guestrin, “Why should i trust you? Explaining the predictions of any classifier,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp. 1135–1144

    2016

  14. [22]

    Cap- tum: A unified and generic model interpretability library for pytorch,

    N. Kokhlikyan, V . Miglani, M. Martin, E. Wang, B. Alsallakh, J. Reynolds, A. Melnikov, N. Kliushkina, C. Araya, S. Yan et al., “Cap- tum: A unified and generic model interpretability library for pytorch,” arXiv preprint arXiv:2009.07896 , 2020

    2009 arXiv

  15. [23]

    ISCAS Benchmark Circuits for Logic Synthesis and Verification,

    J. Hayes, “ISCAS Benchmark Circuits for Logic Synthesis and Verification,” accessed: 2024-11. [Online]. Available: https://web.eecs. umich.edu/∼jhayes/iscas.restore/benchmark.html

    2024

  16. [24]

    The EPFL Combinational Benchmark Suite,

    EPFL-LSI, “The EPFL Combinational Benchmark Suite,” accessed: 2024-11. [Online]. Available: https://www.epfl.ch/labs/lsi/page-102566- en-html/benchmarks/

    2024

  17. [25]

    CEP: Common Evaluation Platform,

    MIT-LL, “CEP: Common Evaluation Platform,” accessed: 2024-11. [Online]. Available: https://github.com/mit-ll/CEP/

    2024

  18. [26]

    SMOTE: synthetic minority over-sampling technique,

    N. V . Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “SMOTE: synthetic minority over-sampling technique,” Journal of arti- ficial intelligence research, vol. 16, pp. 321–357, 2002

    2002

  19. [2014]

    Available: https://eprint.iacr.org/2014/025

    [Online]. Available: https://eprint.iacr.org/2014/025

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.