Behavioral Privacy Leakage in Agent Negotiations: Formalizing and Mitigating Inference Attacks via Randomized Policies

The adoption of autonomous negotiation agents is increasing in high-stakes environments such as insurance and procurement. While existing cryptographic techniques can protect explicitly disclosed constraints, they remain vulnerable to attacks that infer private constraints through behavioral dynamics such as concession trajectories, timing, and convergence patterns.

This study addresses behavioral differential privacy in multi-round negotiation protocols. The researchers designed an adaptive stochastic negotiation policy that simultaneously guarantees the following three properties:

• (ε,δ)-differential privacy guarantee
• Almost-sure convergence of the offer sequence when the opponent's reservation value is admissible
• Maintaining high negotiation utility

Evaluating performance through 3,000 synthetic interest-rate negotiations, the approach maintained negotiation success rate and utility above 90% while reducing the attacker's inference accuracy by 43-50%. This demonstrates that strong privacy guarantees are achievable with minimal performance degradation.