In-Context Unlearning: Language Models as Few Shot Unlearners (2310.07579v4)

Abstract: Machine unlearning, the study of efficiently removing the impact of specific training instances on a model, has garnered increased attention in recent years due to regulatory guidelines such as the \emph{Right to be Forgotten}. Achieving precise unlearning typically involves fully retraining the model and is computationally infeasible in case of very large models such as LLMs. To this end, recent work has proposed several algorithms which approximate the removal of training data without retraining the model. These algorithms crucially rely on access to the model parameters in order to update them, an assumption that may not hold in practice due to computational constraints or having only query access to the LLMs. In this work, we propose a new class of unlearning methods for LLMs called ``In-Context Unlearning.'' This method unlearns instances from the model by simply providing specific kinds of inputs in context, without the need to update model parameters. To unlearn specific training instances, we present these instances to the LLMs at inference time along with labels that differ from their ground truth. Our experimental results demonstrate that in-context unlearning performs on par with, or in some cases outperforms other state-of-the-art methods that require access to model parameters, effectively removing the influence of specific instances on the model while preserving test accuracy.

In-Context Unlearning: LLMs as Few-Shot Unlearners

The paper under review introduces a novel approach to machine unlearning, specifically targeting LLMs through a process termed In-Context Unlearning (ICUL). This methodological innovation addresses key challenges in the current landscape of LLM deployment, such as privacy concerns mandated by regulations like GDPR's "Right to be Forgotten" and issues of copyright infringement. The proposed ICUL method circumvents the need for parameter access or extensive retraining, offering a practical alternative for unlearning without computational burdens.

Methodology and Results

In-Context Unlearning represents a significant shift from traditional machine unlearning paradigms that require direct manipulation of model parameters. Instead, ICUL leverages the inherent capabilities of in-context learning (ICL) by using specifically constructed input contexts at inference time to effectively erase knowledge concerning specific data points. This process involves presenting the targeted data point with a reversed label, along with additional correctly labeled examples, thereby altering the model's output space to mimic one where the data point was never part of the training set.

The authors conduct empirical evaluations using established benchmark datasets like Yelp, SST-2, and Amazon reviews. The experimental results reveal that ICUL achieves competitive unlearning performance compared to existing gradient ascent-based methods that require parameter modification. Specifically, ICUL effectively removes the influence of the targeted data points from the model's outputs, maintaining classification accuracy on unseen test data that rivals or surpasses state-of-the-art methods.

Key Findings

• Black-box Capability: ICUL functions without the need for parameter access, allowing its deployment in scenarios where LLMs are utilized as black-boxes, such as through APIs. This enhances its applicability across diverse industrial and research settings where model introspection is restricted.
• Empirical Efficacy: The paper employs membership inference attacks (MIAs) to validate unlearning success, using the LiRA-Forget metric to demonstrate that unlearned models' outputs on removed data points become indistinguishable from untrained models. This positions ICUL as a robust method for unlearning that does not compromise model efficacy.
• Performance Consistency: Across multiple datasets and model sizes (560 million and 1.1 billion parameters), ICUL consistently reduces the likelihood of MIAs detecting unlearned data points, whilst maintaining competitive classification accuracies.

Implications and Future Directions

The research posits critical implications for the fields of privacy preservation and copyright compliance, with ICUL offering an efficient mechanism to adhere to data deletion requests without necessitating the expensive process of model retraining. This approach aligns with contemporary regulatory frameworks, offering a scalable solution to the challenges of machine unlearning.

Theoretically, the work extends the utility of in-context learning, showcasing its applicability in tasks beyond performance enhancement, such as data removal during inference. The integration of label flipping highlights the nuanced ways in which model behavior can be shaped at inference without direct retraining.

Future research could explore scaling ICUL to larger datasets and more complex tasks, examining the potential for simultaneous unlearning of multiple data points. Additionally, extending this framework to other types of generative models or different learning paradigms may provide broader applications for this unlearning technique.

In sum, this paper provides a comprehensive analysis of a pioneering approach to machine unlearning, delivering insights of both practical and theoretical value to the wider AI and machine learning community.