Title: Quiet Decisions: Copycat-Resilient Best-Arm Identification Speaker: Asaf Cohen (Ben-Gurion University of the Negev)

Speaker: Asaf Cohen (Ben-Gurion University of the Negev) Date: Tuesday, January 20, 2026 Time: 14:00 Location: Room 202, Building 37 Title: Quiet Decisions: Copycat-Resilient Best-Arm Identification Abstract: Information-theoretic secrecy measures such as equivocation, divergence-based detectability and covert capacity have been known for decades in secure communication, yet rarely shaped mainstream practice. In learning and decision-making, they may come back to life. This talk builds a toolkit for Learning Under Watch: how to act and learn while remaining statistically indistinguishable from normal behavior. The focus is best-arm identification in stochastic linear bandits. An agent chooses among K arms with d-dimensional features; each pull returns an inner-product reward plus independent noise. After T pulls the agent must name the best arm. However, a passive observer (“copycat Chloe”) watches the actions, and the aim is also to keep Chloe ignorant of which arm is best. A minimax-optimal strategy attains an error exponent Omega(T/log d) but leaks its estimate because superior arms are sampled more often. A naively secure uniform strategy hides intent yet falls to Omega(T/d). We present Coded Arms: a keyless scheduling scheme that shapes observable play frequencies to mask preference while preserving identification power. Coded Arms achieves Omega(T/log^2 d) and reveals negligible information about the best arm’s identity, offering practical guidance on how to trade off learning speed and stealth. For context, we briefly sketch two related vignettes: Transition-Preserving Attacks in finite MDPs, where perfect covertness is possible when transition statistics match a nominal policy, and detectability limits in linear systems, where certain interventions leave an unavoidable footprint. A unifying rule emerges: match the watcher’s sufficient statistics, via coded exploration and action shaping, while tracking privacy and stealth from equivocation metrics to detection error exponents.