Seminar M.Sc Students
Adversarial Example Generation via Reinforcement Learning
Tomer Zemelman, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisor: Dr. Mahmood Sharif
Advisor: Prof. Noam Koenigstein
Abstract:
Adversarial attacks expose weaknesses in modern machine-learning models by searching for small, human-imperceptible perturbations that cause misclassification. Uncovering and improving such attacks is critical for enhancing model reliability prior to deploying it in safety- and security-critical settings. To this end, this research investigates how reinforcement learning (RL) can help optimize adversarial attacks. We train an RL policy to guide the attack process under a fixed iteration budget, aiming to improve effectiveness and efficiency over common manually tuned procedures. Beyond raw success rates, we introduce an evaluation methodology that estimates per-sample difficulty (conditional on the existence of a feasible adversarial example within the constraint set) and provides a formal way to compare attack methods. We apply our method to enhance a state-of-the-art iterative attacks across multiple architectures and datasets and evaluate both naturally trained and adversarially robust models. We find that, compared to established methods, the RL-optimized attack achieves higher success on difficult samples and reaches successful perturbations in fewer iterations, demonstrating better efficiency (i.e., superior use of the available budget) while maintaining competitive performance on easier samples. These results indicate that learning to attack—casting the attack procedure as a sequential decision problem—can yield practical gains over carefully tuned heuristics.
BIO:
Tomer Zemelman is an M.Sc student in the Department of Industrial Engineering at Tel Aviv University, specializing in Data Science.
He holds a B.Sc. in Industrial Engineering from the Technion—Israel Institute of Technology and works as a Senior Machine Learning Engineer.His research is supervised by Prof. Noam Koenigstein and Dr. Mahmood Sharif, focuses on developing reinforcement learning based optimization of iterative adversarial attacks to reveal worst-case behaviors and support the deployment of robust models in safety and security critical settings.
Bootstrap Stacking for Small Validation Sets
Moti Molochny, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisor: Prof. Amichai Painsky
Abstract:
The efficacy of stacked generalization diminishes rapidly in low-data regimes, where small validation sets cause meta-learner overfitting and weight instability. To address this, we present BootStackr, a novel approach that integrates bootstrap aggregation into the stacking procedure. By averaging weights over resampled validation sets, BootStackr mitigates variance and yields robust ensemble configurations. Our results confirm that BootStackr significantly improves predictive performance over traditional methods whenever validation data is limited.
BIO:
Moti Molochny is an M.Sc. student in the Department of Industrial Engineering at Tel Aviv University, specializing in Data Science. He previously completed a B.Sc. in Industrial Engineering at the same institution. His research centers on model selection within the field of statistical learning.
Daily life impact and detection of migraine through integrated wearable and self-report data: a two-year prospective cohort study
Ilan Vasilevsky, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisor: Prof. Dan Yamin
Abstract:
Migraine is a highly prevalent and disabling neurological disorder that frequently goes undiagnosed. To address gaps in detection and characterization, we conducted a two-year prospective cohort study integrating subjective reports with objective wearable data. We analyzed 750K daily questionnaires and 1.5 million days of smartwatch data from 4,000 participants to investigate the behavioral and biometric signatures of migraine. Results showed that individuals with migraines experienced higher headache frequency characterized by distinct symptom clusters and a uniform annual distribution, differing significantly from the seasonal, respiratory-linked headaches observed in non-migraineurs. Although migraineurs bore a higher overall cumulative burden, the acute physiological impact during headache days was more pronounced in non-migraineurs, suggesting different contexts such as acute infection versus the chronic, adapted nature of migraine. Leveraging these insights, we developed a machine learning model that distinguished migraineurs from controls with an AUC of 0.82. These findings demonstrate that combining digital data with wearable biosensors offers a viable, scalable pathway toward automated migraine screening and improved clinical management.
BIO:
Ilan Vasilevsky, M.Sc. student in the School of Industrial & Intelligent Systems Engineering, Data Scientist Researcher at the Laboratory for Digital Epidemics and Health Analysis (DEHA Lab), actively engaged in a study focused on the detection and daily impact of migraine. His research utilizes physiological data extracted from smartwatches, as well as behavioral information gathered through daily self-reports.
