Apple paper proposes randomized policies to mitigate behavioral privacy leakage in autonomous negotiation agents

Apple’s Machine Learning Research group published a paper accepted at the AI4TCI workshop at the International Conference on Availability, Reliability and Security (ARES) 2026. The work formalizes behavioral privacy leakage in autonomous negotiation agents, a subtler threat than cryptographic protections that focus on explicitly disclosed constraint values. Adversaries can infer private constraints from observable negotiation dynamics such as concession trajectories, timing, and convergence patterns.

The researchers propose an adaptive stochastic negotiation policy designed to jointly guarantee (ε,δ)-differential privacy, ensure almost-sure convergence of the offer sequence when the counterparty’s reservation value permits, and maintain high negotiation utility. In experiments on 3,000 synthetic bilateral negotiations, the mechanism reduced adversarial inference accuracy by 43–50% while preserving negotiation success rates and utility above 90%.

The paper highlights the increasing deployment of autonomous negotiation agents in high-stakes settings such as insurance and procurement, where traditional cryptographic methods are insufficient to address behavioral privacy risks. The proposed approach demonstrates that strong privacy guarantees can be achieved in multi-round negotiation protocols without significant loss of performance.