Machine unlearning is the process of surgically removing the influence of specific training data from a pre-trained model. Driven by privacy regulations like the GDPR’s "Right to be Forgotten," it offers a way to purge sensitive information or copyrighted concepts without the prohibitive cost of retraining from scratch.
However, recent research reveals a hidden trade-off: the process of forgetting often compromises the model's fairness. When a model "unlearns," it doesn't just lose information; it undergoes a distribution shift that can inadvertently amplify biases or degrade performance for specific demographic groups.
The Mechanism of Fairness Degradation
The core intuition behind unlearning-induced bias lies in the non-uniformity of deletion requests. In a perfect world, data removal requests would be distributed randomly across a dataset. As Dawen Zhang, Lina Yao, and their colleagues demonstrate in To be forgotten or to be fair: unveiling fairness implications of machine unlearning methods, random deletions are generally safe.
The problem arises when deletion requests cluster within specific sub-populations. If a particular demographic group exercises their privacy rights more frequently, the remaining training set becomes unrepresentative. This demographic skewing causes the model's loss landscape to shift unevenly, often leading to a "fairness penalty" where the model becomes less accurate or more discriminatory toward the very groups that are already underrepresented. Thomas De Min and Massimiliano Mancini frame this in Group-robust Machine Unlearning as a distribution shift problem, where standard unlearning techniques fail because they assume uniform data removal.
The Efficiency-Fairness Trade-off
To make unlearning computationally feasible, researchers use two primary strategies: Exact Unlearning (via data sharding) and Approximate Unlearning (via weight updates). Both introduce unique fairness challenges.
Data Sharding and the SISA Bottleneck
Exact unlearning often uses the SISA (Sharded, Isolated, Sliced, and Aggregated) framework, which partitions data into independent shards. When a deletion request arrives, only the affected shard is retrained. While efficient, Chaochao Chen and Jiaming Zhang show in the CURE4Rec benchmark that sharding can inadvertently cluster active or similar users together. This results in a scenario where removing data from one shard disproportionately degrades the service quality for specific groups, widening the performance gap between high-activity and low-activity users.
Approximate Updates and Latent Bias
Approximate methods, such as those using influence functions or weight saliency, attempt to "fine-tune" the model away from the target data. Chongyu Fan and Eric Wong propose SalUn in SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation to identify and update only the most "salient" weights. While faster, these shortcuts lack the precision of retraining. In generative models, Yuta Nakashima notes in EMMA that erasing specific concepts can shift the model's latent space toward more homogeneous representations, reinforcing stereotypes by removing the "diverse" features that were entangled with the erased concept.
Proactive and Robust Unlearning
Recognizing these risks, a new direction of "Fair Machine Unlearning" has emerged. The goal is to integrate fairness as a primary constraint rather than an afterthought.
- Fairness Regularization: Alex Oesterling and Himabindu Lakkaraju introduced a framework in Fair Machine Unlearning: Data Removal while Mitigating Disparities that uses second-order updates to maintain fairness guarantees during data removal. By re-balancing the influence of entire demographic groups, they ensure that removing a single sample doesn't skew the model's overall equity.
- Anticipatory Selection: Instead of reacting to deletions, Rachael Hwee Ling Sim and Bryan Kian Hsiang Low propose DADS in Deletion-Anticipative Data Selection with a Limited Budget. This method proactively selects training data that is robust to future unknown deletions, though they caution that even this can introduce bias if certain groups are more likely to request deletion.
- Bi-level Debiasing: In graph-based models, where data points (nodes) are interconnected, Yezi Liu and Yang Ni developed FGU (Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing). This approach uses a bi-level strategy to ensure that the structural changes caused by node removal do not strengthen correlations between model outputs and sensitive attributes like race or gender.
The Adversarial Frontier
Finally, the sensitivity of unlearning to data distribution makes it a target for manipulation. Chenxu Zhao and Mengdi Huai have shown in Static and Sequential Malicious Attacks in the Context of Selective Forgetting that an adversary can use "malicious forgetting requests" to intentionally degrade a model's fairness. By strategically requesting the removal of data from minority groups, an attacker can widen fairness gaps and steer the model toward a biased state, highlighting that unlearning is not just a privacy tool, but a potential security vulnerability.