Efficient Detection and Quantification of Timing Leaks with Neural Networks
Pith reviewed 2026-05-24 17:08 UTC · model grok-4.3
The pith
Neural networks learn timing behaviors of programs with thousands of methods and can be analyzed to quantify timing side-channel leaks.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The authors claim that a neural network can be trained to predict program timing from inputs and that an MILP encoding of this network can then be solved to compute the maximum timing difference between secret-dependent paths, thereby both detecting and quantifying the strength of timing side channels at a scale unreachable by direct static analysis.
What carries the argument
Neural network timing model combined with MILP encoding of the network to estimate side-channel strength.
If this is right
- Real-world applications too large for static analysis can still have their timing leaks detected and measured.
- Leakage is reported as a quantity rather than a binary answer, allowing direct comparison of threat levels.
- The two-task split keeps both the learning step and the MILP step computationally feasible for networks with thousands of neurons.
Where Pith is reading between the lines
- If the same modeling technique generalizes, learned timing models could be reused to check other observable behaviors such as cache or power side channels.
- Quantified leak values could be fed back into compiler passes that search for constant-time rewrites of the most leaky code paths.
- The approach suggests a template for using learned models plus optimization for other dynamic security properties beyond timing.
Load-bearing premise
The neural network's predicted timing must match the program's actual timing closely enough for the MILP results on the model to reflect real leaks.
What would settle it
Executing the original program on inputs that the MILP analysis flags as producing a large timing gap and measuring that the observed times show no such gap would falsify the claim that the model-based quantification is reliable.
Figures
read the original abstract
Detection and quantification of information leaks through timing side channels are important to guarantee confidentiality. Although static analysis remains the prevalent approach for detecting timing side channels, it is computationally challenging for real-world applications. In addition, the detection techniques are usually restricted to 'yes' or 'no' answers. In practice, real-world applications may need to leak information about the secret. Therefore, quantification techniques are necessary to evaluate the resulting threats of information leaks. Since both problems are very difficult or impossible for static analysis techniques, we propose a dynamic analysis method. Our novel approach is to split the problem into two tasks. First, we learn a timing model of the program as a neural network. Second, we analyze the neural network to quantify information leaks. As demonstrated in our experiments, both of these tasks are feasible in practice --- making the approach a significant improvement over the state-of-the-art side channel detectors and quantifiers. Our key technical contributions are (a) a neural network architecture that enables side channel discovery and (b) an MILP-based algorithm to estimate the side-channel strength. On a set of micro-benchmarks and real-world applications, we show that neural network models learn timing behaviors of programs with thousands of methods. We also show that neural networks with thousands of neurons can be efficiently analyzed to detect and quantify information leaks through timing side channels.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes splitting timing side-channel analysis into two stages: (1) learning a neural-network surrogate for a program's timing behavior from dynamic execution traces, and (2) encoding the trained NN as a mixed-integer linear program (MILP) whose solution yields bounds on timing differences over secret inputs, thereby detecting and quantifying leaks. Experiments on micro-benchmarks and real-world applications are reported to show that NNs can be trained for programs containing thousands of methods and that the resulting MILPs remain tractable for networks with thousands of neurons.
Significance. If the NN models prove sufficiently faithful and the MILP bounds are shown to be reliable, the method would supply a practical dynamic route to both detection and quantitative leak assessment in applications where static analysis is intractable. The technical device of an MILP encoding of a trained NN is a concrete contribution that could be reused beyond this setting.
major comments (2)
- [Experimental Evaluation] Experimental Evaluation (and abstract): no quantitative fidelity metrics (MAE, R², or held-out trace error) or direct validation of MILP-derived leak bounds against measurements on the original program are supplied. Without these, the central claim that the two-stage procedure yields sound quantifications rests on an unchecked assumption that the learned NN is a faithful surrogate.
- [Method] Method (two-stage description): the paper does not bound or mitigate systematic approximation error introduced by the NN (e.g., on branches, cache effects, or dispatch) before the MILP optimizer is applied; any such bias can be amplified by the optimizer, producing spurious or missed leaks. A concrete error-propagation argument or empirical sanity check against the original program is required for the quantification results to be load-bearing.
minor comments (1)
- [Abstract] The abstract asserts a 'significant improvement over the state-of-the-art' without naming the baselines or reporting quantitative deltas; a short comparison table would clarify the claim.
Simulated Author's Rebuttal
We thank the referee for the constructive comments, which highlight important aspects for strengthening the presentation of our results. We address each major comment below and will revise the manuscript accordingly.
read point-by-point responses
-
Referee: [Experimental Evaluation] Experimental Evaluation (and abstract): no quantitative fidelity metrics (MAE, R², or held-out trace error) or direct validation of MILP-derived leak bounds against measurements on the original program are supplied. Without these, the central claim that the two-stage procedure yields sound quantifications rests on an unchecked assumption that the learned NN is a faithful surrogate.
Authors: We agree that quantitative fidelity metrics and direct validation of the derived bounds would make the claims more robust. In the revised version we will report MAE, R², and held-out trace prediction error for all trained networks. For the micro-benchmarks we will also add side-by-side comparisons of MILP-computed timing differences against exhaustive measurements on the original programs; for the larger applications we will report the same metrics on a held-out set of secret inputs. revision: yes
-
Referee: [Method] Method (two-stage description): the paper does not bound or mitigate systematic approximation error introduced by the NN (e.g., on branches, cache effects, or dispatch) before the MILP optimizer is applied; any such bias can be amplified by the optimizer, producing spurious or missed leaks. A concrete error-propagation argument or empirical sanity check against the original program is required for the quantification results to be load-bearing.
Authors: We acknowledge that the manuscript does not yet contain an explicit error-propagation analysis. In the revision we will add a dedicated subsection discussing sources of approximation error (branch misprediction, cache effects, method dispatch) and their potential amplification under MILP optimization. We will also include the empirical sanity checks described in the response to the first comment and will state the resulting limitations on the soundness of the quantified bounds. revision: yes
Circularity Check
No circularity in two-stage NN learning and MILP analysis
full rationale
The paper's core derivation splits into independent stages: (1) supervised learning of an NN timing model from dynamic execution traces, and (2) separate MILP encoding and optimization over the fixed NN weights to bound output differences. Neither stage reduces to the other by definition or construction; the MILP step treats the trained NN as an external black-box function rather than re-deriving fitted parameters. No self-citations, uniqueness theorems, or ansatzes are invoked in the provided text to justify the central claims. The approach is therefore self-contained against external benchmarks (trace data and MILP solvers) and receives the default non-circularity finding.
Axiom & Free-Parameter Ledger
Reference graph
Works this paper leans on
-
[1]
American fuzzy lop (2016), http://lcamtuf.coredump.cx/afl/
work page 2016
-
[2]
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., Devin, M., Ghe- mawat, S., Irving, G., Isard, M., et al.: Tensorflow: A system for large-scale machine learning. In: OSDI’16. pp. 265–283 (2016)
work page 2016
-
[3]
In: Proceedings of the 27th ACM SIGPLAN-SIGACT symposium on Principles of programming languages
Agat, J.: Transforming out timing leaks. In: Proceedings of the 27th ACM SIGPLAN-SIGACT symposium on Principles of programming languages. pp. 40–
-
[4]
Almeida, J.B., Barbosa, M., Barthe, G., Dupressoir, F., Emmi, M.: Verify- ing constant-time implementations. In: USENIX Security Symposium. pp. 53–70 (2016)
work page 2016
-
[5]
Antonopoulos, T., Gazzillo, P., Hicks, M., Koskinen, E., Terauchi, T., Wei, S.: De- composition instead of self-composition for proving the absence of timing channels. In: ACM SIGPLAN Notices. vol. 52, pp. 362–375. ACM (2017)
work page 2017
-
[6]
Online: https://github.com/Apogee-Research/STAC
Apogee-Research: Space/time analysis for cybersecurity (stac) repository. Online: https://github.com/Apogee-Research/STAC
-
[7]
Applied Soft Computing 33, 263–277 (2015)
Arar, ¨O.F., Ayan, K.: Software defect prediction using cost-sensitive neural net- work. Applied Soft Computing 33, 263–277 (2015)
work page 2015
-
[8]
Arora, R., Basu, A., Mianjy, P., Mukherjee, A.: Understanding Deep Neural Net- works with Rectified Linear Units. arXiv e-prints (2016)
work page 2016
-
[9]
In: Proceedings of the 17th ACM conference on Computer and commu- nications security
Askarov, A., Zhang, D., Myers, A.C.: Predictive black-box mitigation of timing channels. In: Proceedings of the 17th ACM conference on Computer and commu- nications security. pp. 297–307. ACM (2010)
work page 2010
-
[10]
Backes, M., K¨ opf, B., Rybalchenko, A.: Automatic discovery and quantification of information leaks. In: S&P’09 (2009)
work page 2009
-
[11]
In: Computer Security Foundations Workshop, 2004
Barthe, G., D’Argenio, P.R., Rezk, T.: Secure information flow by self-composition. In: Computer Security Foundations Workshop, 2004. Proceedings. 17th IEEE. pp. 100–114. IEEE (2004)
work page 2004
-
[12]
Computer Networks 48(5), 701–716 (2005)
Brumley, D., Boneh, D.: Remote timing attacks are practical. Computer Networks 48(5), 701–716 (2005)
work page 2005
-
[13]
Chen, J., Feng, Y., Dillig, I.: Precise detection of side-channel vulnerabilities using quantitative cartesian hoare logic. In: CCS (2017)
work page 2017
-
[14]
Chen, S., Wang, R., Wang, X., Zhang, K.: Side-channel leaks in web applications: A reality today, a challenge tomorrow. In: S&P’10 (2010)
work page 2010
-
[15]
Courbariaux, M., Hubara, I., Soudry, D., El-Yaniv, R., Bengio, Y.: Binarized neural networks: Training deep neural networks with weights and activations constrained to+ 1 or-1. arXiv preprint arXiv:1602.02830 (2016)
work page internal anchor Pith review Pith/arXiv arXiv 2016
-
[16]
Mathe- matics of Control, Signals, and Systems 2, 303–314 (12 1989)
Cybenko, G.: Approximation by superpositions of a sigmoidal function. Mathe- matics of Control, Signals, and Systems 2, 303–314 (12 1989)
work page 1989
-
[17]
ACM Transactions on Information and System Security (TISSEC) 18(1), 4 (2015)
Doychev, G., K¨ opf, B., Mauborgne, L., Reineke, J.: Cacheaudit: A tool for the static analysis of cache side channels. ACM Transactions on Information and System Security (TISSEC) 18(1), 4 (2015)
work page 2015
-
[18]
In: NASA Formal Methods Symposium
Dutta, S., Jha, S., Sankaranarayanan, S., Tiwari, A.: Output range analysis for deep feedforward neural networks. In: NASA Formal Methods Symposium. pp. 121–138. Springer (2018)
work page 2018
-
[19]
In: International Conference on Computer Aided Verification
Eldib, H., Wang, C.: Synthesis of masking countermeasures against side channel attacks. In: International Conference on Computer Aided Verification. pp. 114–130. Springer (2014) Quantification of Leaks with NNs 17
work page 2014
-
[20]
ACM Transactions on Software Engineering and Methodology (TOSEM) 24(2), 11 (2014)
Eldib, H., Wang, C., Schaumont, P.: Formal verification of software countermea- sures against side-channel attacks. ACM Transactions on Software Engineering and Methodology (TOSEM) 24(2), 11 (2014)
work page 2014
-
[21]
Constraints 23(3), 296–309 (2018)
Fischetti, M., Jo, J.: Deep neural networks and mixed integer linear optimization. Constraints 23(3), 296–309 (2018)
work page 2018
-
[22]
In: Security and Privacy, 1982 IEEE Symposium on
Goguen, J.A., Meseguer, J.: Security policies and security models. In: Security and Privacy, 1982 IEEE Symposium on. pp. 11–11. IEEE (1982)
work page 1982
-
[23]
Goldsmith, S.F., Aiken, A.S., Wilkerson, D.S.: Measuring empirical computational complexity. In: FSE’07. pp. 395–404. ACM (2007)
work page 2007
-
[24]
Guo, S., Wu, M., Wang, C.: Adversarial symbolic execution for detecting concurrency-related cache timing leaks. In: Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering. pp. 377–388. ACM (2018)
work page 2018
-
[25]
Gurobi Optimization, L.: Gurobi optimizer reference manual (2018), http://www. gurobi.com
work page 2018
-
[26]
Neural Networks 2, 359–366 (1989)
Hornik, K., Stinchcombe, M.B., White, H.: Multilayer feedforward networks are universal approximators. Neural Networks 2, 359–366 (1989)
work page 1989
-
[27]
In: 2013 IEEE Symposium on Security and Privacy
Hund, R., Willems, C., Holz, T.: Practical timing side channel attacks against kernel space aslr. In: 2013 IEEE Symposium on Security and Privacy. pp. 191–205. IEEE (2013)
work page 2013
-
[28]
In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security
Kersten, R., Luckow, K., P˘ as˘ areanu, C.S.: Poster: Afl-based fuzzing for java with kelinci. In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security. pp. 2511–2513. ACM (2017)
work page 2017
-
[29]
Adam: A Method for Stochastic Optimization
Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)
work page internal anchor Pith review Pith/arXiv arXiv 2014
-
[30]
In: Annual International Cryptology Conference
Kocher, P.C.: Timing attacks on implementations of diffie-hellman, rsa, dss, and other systems. In: Annual International Cryptology Conference. pp. 104–113. Springer (1996)
work page 1996
-
[31]
K¨ opf, B., Basin, D.: An information-theoretic model for adaptive side-channel at- tacks. In: CCS’07. pp. 286–296 (2007)
work page 2007
-
[32]
K¨ opf, B., D¨ urmuth, M.: A provably secure and efficient countermeasure against timing attacks. In: CSF’09 (2009)
work page 2009
-
[33]
In: 2017 IEEE/ACM 39th Inter- national Conference on Software Engineering (ICSE)
Landman, D., Serebrenik, A., Vinju, J.J.: Challenges for static analysis of java reflection-literature review and empirical study. In: 2017 IEEE/ACM 39th Inter- national Conference on Software Engineering (ICSE). pp. 507–518. IEEE (2017)
work page 2017
-
[34]
Online post at: https: //rdist.root.org/2009/05/28/timing-attack-in-google-keyczar-library/ (2009)
Lawson, N.: Timing attack in google keyczar library. Online post at: https: //rdist.root.org/2009/05/28/timing-attack-in-google-keyczar-library/ (2009)
work page 2009
-
[35]
LeCun, Y., Bengio, Y., Hinton, G.: Deep learning. nature 521(7553), 436 (2015)
work page 2015
-
[36]
VulDeePecker: A Deep Learning-Based System for Vulnerability Detection
Li, Z., Zou, D., Xu, S., Ou, X., Jin, H., Wang, S., Deng, Z., Zhong, Y.: Vuldeepecker: A deep learning-based system for vulnerability detection. arXiv:1801.01681 (2018)
work page internal anchor Pith review Pith/arXiv arXiv 2018
-
[37]
Livshits, V.B., Lam, M.S.: Finding security vulnerabilities in java applications with static analysis. In: USENIX Security Symposium. vol. 14, pp. 18–18 (2005)
work page 2005
-
[38]
In: Formal Techniques for Distributed Systems, pp
Milushev, D., Beck, W., Clarke, D.: Noninterference via symbolic execution. In: Formal Techniques for Distributed Systems, pp. 152–168. Springer (2012)
work page 2012
-
[39]
Biometrika 78(3), 691–692 (1991)
Nagelkerke, N.J., et al.: A note on a general definition of the coefficient of deter- mination. Biometrika 78(3), 691–692 (1991)
work page 1991
-
[40]
Narodytska, N., Kasiviswanathan, S., Ryzhyk, L., Sagiv, M., Walsh, T.: Verifying properties of binarized deep neural networks. In: AAAI’18 (2018)
work page 2018
-
[41]
DifFuzz: Differential Fuzzing for Side-Channel Analysis
Nilizadeh, S., Noller, Y., Pasareanu, C.S.: Diffuzz: Differential fuzzing for side- channel analysis. ICSE (2019), http://arxiv.org/abs/1811.07005 18 Tizpaz-Niari, et al
work page internal anchor Pith review Pith/arXiv arXiv 2019
-
[42]
Rosner, N., Burak Kadron, I., Bang, L., Bultan, T.: Profit: Detecting and quanti- fying side channels in networked applications. NDSS (2019)
work page 2019
-
[43]
IEEE Jour- nal on selected areas in communications (2003)
Sabelfeld, A., Myers, A.C.: Language-based information-flow security. IEEE Jour- nal on selected areas in communications (2003)
work page 2003
-
[44]
Smith, G.: On the foundations of quantitative information flow. In: FoSSaCS’09 (2009)
work page 2009
-
[45]
Sung, C., Paulsen, B., Wang, C.: Canal: A cache timing analysis framework via llvm transformation. pp. 904–907. ASE 2018
work page 2018
-
[46]
Tang, T.A., Mhamdi, L., McLernon, D., Zaidi, S.A.R., Ghogho, M.: Deep learn- ing approach for network intrusion detection in software defined networking. In: WINCOM’16 (2016)
work page 2016
-
[47]
libFuzzer team, G.: Libfuzzer: Coverage-based fuzz testing. Online: http://llvm. org/docs/LibFuzzer.html (2016)
work page 2016
-
[48]
In: Interna- tional Static Analysis Symposium
Terauchi, T., Aiken, A.: Secure information flow as a safety problem. In: Interna- tional Static Analysis Symposium. pp. 352–367. Springer (2005)
work page 2005
-
[49]
Tizpaz-Niari, S., ˇCern´ y, P., Chang, B.Y.E., Sankaranarayanan, S., Trivedi, A.: Discriminating traces with time. In: TACAS’17. pp. 21–37. Springer (2017)
work page 2017
-
[50]
Tizpaz-Niari, S., ˇCern´ y, P., Chang, B.E., Trivedi, A.: Differential performance debugging with discriminant regression trees. In: AAAI’18. pp. 2468–2475 (2018)
work page 2018
-
[51]
arXiv preprint arXiv:1808.10502 (2018)
Tizpaz-Niari, S., ˇCern´ y, P., Trivedi, A.: Data-driven debugging for functional side channels. arXiv:1808.10502 (2018)
-
[52]
In: Computer Aided Verification (CAV)
Tizpaz-Niari, S., ˇCern´ y, P., Trivedi, A.: Quantitative mitigation of timing side channels. In: Computer Aided Verification (CAV). pp. 140–160 (2019)
work page 2019
-
[53]
Mitigating Power Side Channels during Compilation
Wang, J., Sung, C., Wang, C.: Mitigating power side channels during compilation. arXiv preprint arXiv:1902.09099 (2019)
work page internal anchor Pith review Pith/arXiv arXiv 1902
-
[54]
In: 26th USENIX Security Symposium
Wang, S., Wang, P., Liu, X., Zhang, D., Wu, D.: Cached: Identifying cache-based timing channels in production software. In: 26th USENIX Security Symposium. pp. 235–252 (2017)
work page 2017
-
[55]
In: Proceedings of the 27th ACM SIGSOFT International Symposium on Software Testing and Analysis
Wu, M., Guo, S., Schaumont, P., Wang, C.: Eliminating timing side-channel leaks using program repair. In: Proceedings of the 27th ACM SIGSOFT International Symposium on Software Testing and Analysis. pp. 15–26. ACM (2018)
work page 2018
-
[56]
In: Pro- ceedings of the 17th ACM conference on Computer and communications security
Zhang, K., Li, Z., Wang, R., Wang, X., Chen, S.: Sidebuster: automated detection and quantification of side-channel leaks in web application development. In: Pro- ceedings of the 17th ACM conference on Computer and communications security. pp. 595–606. ACM (2010)
work page 2010
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.