A Survey of Machine Unlearning (2209.02299v6)

Abstract: Today, computer systems hold large amounts of personal data. Yet while such an abundance of data allows breakthroughs in artificial intelligence, and especially ML, its existence can be a threat to user privacy, and it can weaken the bonds of trust between humans and AI. Recent regulations now require that, on request, private information about a user must be removed from both computer systems and from ML models, i.e. `the right to be forgotten''). While removing data from back-end databases should be straightforward, it is not sufficient in the AI context as ML models oftenremember' the old data. Contemporary adversarial attacks on trained models have proven that we can learn whether an instance or an attribute belonged to the training data. This phenomenon calls for a new paradigm, namely machine unlearning, to make ML models forget about particular data. It turns out that recent works on machine unlearning have not been able to completely solve the problem due to the lack of common frameworks and resources. Therefore, this paper aspires to present a comprehensive examination of machine unlearning's concepts, scenarios, methods, and applications. Specifically, as a category collection of cutting-edge studies, the intention behind this article is to serve as a comprehensive resource for researchers and practitioners seeking an introduction to machine unlearning and its formulations, design criteria, removal requests, algorithms, and applications. In addition, we aim to highlight the key findings, current trends, and new research areas that have not yet featured the use of machine unlearning but could benefit greatly from it. We hope this survey serves as a valuable resource for ML researchers and those seeking to innovate privacy technologies. Our resources are publicly available at https://github.com/tamlhp/awesome-machine-unlearning.

A Survey of Machine Unlearning

The concept of machine unlearning has emerged as a response to privacy regulations demanding the erasure of individual data from computational systems. The paper "A Survey of Machine Unlearning" provides an extensive examination of methods, scenarios, applications, and existing challenges related to machine unlearning, serving as a crucial resource for researchers interested in this field.

Key Insights

1. Definitions and Frameworks:

The paper categorizes unlearning into exact and approximate unlearning. Exact unlearning achieves complete symmetry between a model trained without certain data and a model with data removed. In contrast, approximate unlearning allows some level of residual information in exchange for computational efficiency. The framework emphasizes that while retraining ensures perfect unlearning, it's often computationally prohibitive, leading to the exploration of alternative methods like influence functions and statistical queries.

2. Unlearning Scenarios:

The scenarios explored include zero-glance, zero-shot, and few-shot unlearning, defined by the availability and characteristics of the data to be forgotten. These settings present varying levels of challenges, demanding specialized techniques for effective unlearning.

3. Algorithms and Techniques:

The paper classifies existing approaches into three categories: model-agnostic, model-intrinsic, and data-driven. Model-agnostic methods, such as certified removal, offer general applications but may lack the efficiency of model-specific approaches like those designed for deep neural networks. Data-driven approaches leverage strategies like data partitioning and augmentation to facilitate unlearning.

4. Applications and Challenges:

Beyond compliance with privacy regulations, the paper highlights practical applications in enhancing model robustness against adversarial attacks and correcting biases. Federated learning presents unique challenges due to decentralized data storage, requiring innovations like efficient client removal methodologies.

5. Verification and Metrics:

Verification of unlearning processes is crucial for compliance and trust. Metrics such as activation distances and membership inference attacks assess the effectiveness of unlearning algorithms. However, the paper notes a lack of standardized benchmarking across different methodologies.

6. Future Directions:

Several open questions remain, such as the development of unified design requirements and benchmarking for comparative evaluations. The authors call for advances in adversarial unlearning, interpretable unlearning processes, and further integration of causality to ensure comprehensive data removal.

Implications and Future Research

The implications of machine unlearning are significant across various domains where data privacy is paramount. Future research could address the development of interpretable machine unlearning methods, potentially enhancing confidence in AI systems through transparent unlearning processes. Additionally, as machine learning and data architectures evolve, continuous adaptation and integration of unlearning methodologies will be essential.

In conclusion, this survey acts as a foundational resource for advancing machine unlearning, summarizing current methodologies, challenges, and providing direction for future advancements in data privacy technologies.