Conflict-Aware Machine Unlearning via Kalai–Smorodinsky Bargaining

Modern generative models are continuously updated and adapted rather than remaining static after training, raising increasing concerns about privacy leakage and the unintended retention of sensitive information. Machine unlearning addresses this challenge by removing the influence of targeted concepts or data from a trained model while preserving its overall utility. In generative models, this objective can be formulated as a multi-objective optimization problem that jointly seeks effective forgetting and utility retention. These objectives are inherently competing, and conventional optimization strategies may favor one objective over the other, resulting in either insufficient forgetting or substantial degradation in model performance. We propose Conflict-Aware 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.