Machine Unlearning: A Comprehensive Survey (2405.07406v2)

Abstract: As the right to be forgotten has been legislated worldwide, many studies attempt to design unlearning mechanisms to protect users' privacy when they want to leave machine learning service platforms. Specifically, machine unlearning is to make a trained model to remove the contribution of an erased subset of the training dataset. This survey aims to systematically classify a wide range of machine unlearning and discuss their differences, connections and open problems. We categorize current unlearning methods into four scenarios: centralized unlearning, distributed and irregular data unlearning, unlearning verification, and privacy and security issues in unlearning. Since centralized unlearning is the primary domain, we use two parts to introduce: firstly, we classify centralized unlearning into exact unlearning and approximate unlearning; secondly, we offer a detailed introduction to the techniques of these methods. Besides the centralized unlearning, we notice some studies about distributed and irregular data unlearning and introduce federated unlearning and graph unlearning as the two representative directions. After introducing unlearning methods, we review studies about unlearning verification. Moreover, we consider the privacy and security issues essential in machine unlearning and organize the latest related literature. Finally, we discuss the challenges of various unlearning scenarios and address the potential research directions.

Overview of "Machine Unlearning: A Comprehensive Survey"

The paper "Machine Unlearning: A Comprehensive Survey" by Wang, Tian, and Yu provides a thorough examination of the domain of machine unlearning, driven by the legislative requirements around the right to be forgotten. As data privacy becomes paramount in the field of ML, machine unlearning aims to allow a trained model to remove the influence of specific data points that users wish to have forgotten. This survey categorizes and assesses current methodologies in machine unlearning, framing their distinctions and identifying unresolved issues in the field.

Classification of Machine Unlearning Methods

The survey classifies existing unlearning approaches into four primary scenarios:

1. Centralized Unlearning: It serves as the foundation for most machine unlearning methods and is further subdivided into exact unlearning and approximate unlearning. Exact unlearning generally employs sharding or ensembling techniques to update only relevant sections of the model affected by the data removal request. This method, however, requires substantial storage to retrain efficiently. Conversely, approximate unlearning deals with estimating the data's contribution and aims to remove it without retraining the entire model. It allows for less storage use but may lead to substantial accuracy degradation if not properly bounded.
2. Distributed and Irregular Data Unlearning: This includes federated unlearning, which focuses on distributed federated learning settings, and graph unlearning, which deals with the complex structures inherent in graph data. Here, data segmentation and sparse updates to the nodes and edges play a crucial role.
3. Unlearning Verification: Evaluating the effectiveness of unlearning methods is critical. The paper covers several approaches for verification, such as L2-norm, Kullback-Leibler divergence, and membership inference attacks, which reveal how closely an unlearned model matches a retrained model from scratch.
4. Privacy and Security Concerns: Despite unlearning being a privacy-preserving mechanism, it is susceptible to new security threats like membership inference and information reconstruction attacks. The paper posits that handling such threats necessitates even more robust unlearning and privacy-preservation techniques.

Implications and Future Directions

The survey highlights significant numerical results where certain unlearning strategies outperform naive retraining in computational efficiency without sacrificing model accuracy. It calls attention to the novel applications of machine unlearning in counteracting anomalies, such as backdoor attacks and data anomalies. There is also acknowledgment of privacy risks inherent in the updates made to models through unlearning requests.

The implications of this work are profound in both practical and theoretical contexts. In practice, the survey provides a roadmap for industries seeking to comply with privacy laws efficiently. Theoretically, it opens avenues for future advancements in designing more efficient and reliable unlearning mechanisms that can work seamlessly without a significant hit on the model's performance or storage requirements.

Speculation on Future Developments in AI

As discussed in this survey, future AI systems will likely incorporate unlearning as a standard capability, essential for maintaining compliance with evolving privacy standards. This will involve breakthroughs in mitigating the inherent stochasticity and incrementality challenges of training complex ML models. Future work may also delve into automating these processes, enabling more dynamic and real-time unlearning capabilities.

Machine unlearning is anticipated to grow in importance as AI systems integrate increasingly into personal and sensitive areas of human life—potentially leading to developments where AI adapts its behavior by forgetting irrelevant or outdated information, aligned with user preferences and privacy concerns.

Conclusion

The comprehensive survey by Wang, Tian, and Yu sets a critical stage in the exploration and establishment of machine unlearning. It underscores the non-trivial challenges and highlights the balance needed between efficiency, efficacy, and privacy. The research community is prompted to address open questions regarding stochasticity and the catastrophic nature of unlearning—ensuring models can forget specific information without collateral loss to their operational capabilities.