MUKSB: Machine Unlearning via Kalai–Smorodinsky Bargaining

Modern generative models are continuously updated and adapted rather than remaining fixed after training, which introduces growing concerns around privacy leakage and unintended exposure of sensitive information. Machine unlearning aims to eliminate the influence of such concepts from a trained model while maintaining its overall utility. In generative models, this can be formulated as a dual-objective optimization problem that balances forgetting and retention, but these objectives often conflict, and one can dominate the optimization process, leading to ineffective unlearning or degraded performance.

We propose MUKSB (Machine Unlearning via Kalai–Smorodinsky Bargaining), which models the interplay between forgetting and retention as a coordinated optimization process inspired by game theory, allowing both objectives to contribute proportionally while mitigating imbalance during updates. We evaluate our methodology on image classification and text-to-image generation benchmarks, including concept removal in diffusion models and unlearning in image editing tasks. Our approach achieves stronger forgetting with minimal utility loss and demonstrates improved robustness to adversarial prompts while resisting relearning of distinct object-level concepts, thus maintaining high generation image quality.