Causal Unlearning in Collaborative Optimization: Exact and Approximate Influence Reversal under Adversarial Contributions
The paper presents a method called HF-KCU for causal unlearning in federated learning systems. This method allows for the removal of a client's contribution while maintaining model accuracy and efficiency. It achieves significant speedup in retraining processes and ensures that only affected clients receive updates, thereby preserving the quality of the model for others.
- ▪HF-KCU reduces the complexity of influence removal from O(d^3) to O(kd), where k is much smaller than d.
- ▪The method provides a 47.75 times speedup over traditional retraining while maintaining test accuracy within 0.60% of the baseline.
- ▪Membership inference attacks on the forget set yield success rates matching the retrained model, confirming effective privacy restoration.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20341 (cs) [Submitted on 19 May 2026] Title:Causal Unlearning in Collaborative Optimization: Exact and Approximate Influence Reversal under Adversarial Contributions Authors:Ali Mahdavi, Azadeh Zamanifar, Amirfarhad Farhadi, Omid Kashefi View a PDF of the paper titled Causal Unlearning in Collaborative Optimization: Exact and Approximate Influence Reversal under Adversarial Contributions, by Ali Mahdavi and 3 other authors View PDF HTML (experimental) Abstract:Federated learning systems must support data deletion requests to comply with privacy regulations, yet retraining from scratch after each deletion is computationally prohibitive.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
