Behavioral Privacy Leakage in Agentic Negotiation: Formalizing and Mitigating Inference Attacks via Randomized Policies

This paper studies behavioral privacy leakage in multi-round negotiation agents. Even when constraint values remain hidden, an adversary may infer them from concession trajectories, response timing, and convergence behavior. The authors propose an adaptive stochastic negotiation policy intended to provide (ε, δ)-differential privacy over observable behavior, almost-sure offer-sequence convergence, and high negotiation utility. In 3,000 synthetic bilateral negotiations, the abstract reports a 43–50% reduction in adversarial inference accuracy while negotiation success rate and utility remained above 90%. The abstract does not state the specific privacy parameters, baselines, or evaluation details.