Automation and machine learning are reshaping both the attack and defense landscape. Our work empirically measures AI-driven threats in the wild — from adversarial attacks on defensive classifiers to large-scale CAPTCHA-solving — and builds uncertainty-aware, multi-modal systems that improve defender agility without sacrificing reliability.
We conducted a large-scale empirical measurement of CAPTCHA security across the web, analyzing the gap between the assurances CAPTCHA providers advertise and their practical resistance to automated solvers. The study reveals systematic weaknesses in deployed CAPTCHA schemes and quantifies the conditions under which automated adversaries can bypass them reliably.
Security classifiers trained on historical attack data are brittle targets: adversaries can craft inputs that evade detection while preserving malicious functionality. This work studies adversarial perturbations against real-world security models, measuring both attack effectiveness and the coverage gaps that make multi-modal defenses necessary.
EnSolver investigates how ensemble-based uncertainty quantification can improve the reliability of automated solvers under distribution shift. Rather than maximizing raw accuracy, the system reasons about its own confidence — abstaining when evidence is ambiguous — making it a more rigorous testbed for evaluating the robustness of challenge-response defenses.
← Back to Research Full publication list