Machine Unlearning in Generative AI: A Survey

🌈 Abstract

The article provides a comprehensive survey on machine unlearning (MU) techniques for generative AI (GenAI) models, including large language models (LLMs), generative image models, and multimodal (large) language models (M(L)LMs). It highlights the key differences between MU for traditional AI and GenAI, and formulates three main objectives for effective GenAI unlearning: Accuracy, Locality, and Generalizability. The article categorizes existing unlearning techniques into two main approaches: Parameter Optimization and In-Context Unlearning, and discusses the advantages and limitations of each. It also covers commonly used datasets and benchmarks for evaluating GenAI unlearning, as well as various real-world applications such as safety alignment, privacy compliance, copyright protection, hallucination reduction, and bias alleviation. Finally, the article discusses the challenges and future research directions in this rapidly evolving field.

🙋 Q&A

[01] Objectives of GenAI Unlearning

1. What are the three main objectives for effective GenAI unlearning? The three main objectives for effective GenAI unlearning are:

• Accuracy: The unlearned model should not generate the data points in the seen forget set.
• Locality: The unlearned model should maintain its performance on the retain set.
• Generalizability: The model should generalize the unlearning to the unseen forget set.

2. How do these objectives differ from traditional machine unlearning? The key difference is that the forget set in GenAI can be defined as or simplified as , where can be any undesired model output, including leaked privacy, harmful information, bias and discrimination, copyright data, etc., and is anything that prompts the model to generate . This is in contrast to traditional machine unlearning, which mainly focused on tuning a model to be the same as trained only on .

[02] Categorization of GenAI Unlearning Techniques

1. What are the two main categories of GenAI unlearning techniques discussed in the article? The two main categories of GenAI unlearning techniques are:

1. Parameter Optimization
2. In-Context Unlearning

2. Can you briefly explain the key differences between these two categories? Parameter Optimization techniques focus on adjusting specific model parameters to selectively unlearn certain behaviors without affecting other functions. In contrast, In-Context Unlearning techniques retain the parameters in their original state and manipulate the model's context or environment to facilitate unlearning, without directly modifying the model's internal parameters.

3. What are the advantages and limitations of each category? Parameter Optimization techniques can be effective for unlearning accuracy and generalizability, but may negatively impact model locality. In-Context Unlearning is resource-efficient but only modifies the model's immediate outputs without fundamentally eradicating the unwanted knowledge embedded within the model's parameters.

[03] Datasets and Benchmarks

1. What are some of the key datasets used for evaluating GenAI unlearning? Some of the key datasets used for evaluating GenAI unlearning include:

• Civil Comments, PKU-SafeRLHF, and LAION for safety alignment
• Harry Potter, BookCorpus, and TOFU for copyright protection
• HaluEVAL, CounterFact, and TruthfulQA for hallucination reduction
• The Pile, Yelp/Amazon Reviews, and SST-2 for privacy compliance
• StereoSet and CrowS Pairs for bias/unfairness alleviation

2. Can you provide examples of benchmarks used to evaluate GenAI unlearning? Some examples of benchmarks used to evaluate GenAI unlearning include:

• UnlearnCanvas for evaluating how diffusion models can forget certain styles or objects
• TOFU for evaluating the unlearning capabilities of large language models
• WMDP for identifying and mitigating the risk of LLMs aiding in the development of dangerous weapons
• Object HalBench, MMHal-Bench, and POPE for evaluating hallucinations in multimodal (large) language models

[04] Applications of GenAI Unlearning

1. What are some of the key applications of GenAI unlearning techniques? The key applications of GenAI unlearning techniques include:

• Safety alignment: Reducing the generation of inappropriate, harmful, or illegal content
• Privacy compliance: Removing sensitive or personal information from the model
• Copyright protection: Preventing the generation of copyrighted content
• Hallucination reduction: Eliminating the generation of false or inaccurate information
• Bias alleviation: Reducing biases and unfairness in the model's outputs

2. How do GenAI unlearning techniques compare to other approaches like RLHF for addressing these applications? GenAI unlearning techniques offer several advantages over RLHF:

• Unlearning requires only the collection of target negative examples and unseen samples, which is easier and cheaper than the positive samples needed for RLHF.
• Unlearning is computationally efficient, as it can usually be accomplished with the same resources required for fine-tuning, unlike the extensive training required for RLHF.