Neuralyzing the Trace: Selective Representation-Level Unlearning with Contrastive Sparse Autoencoders

As pretrained models are increasingly shared on the web, ensuring that models can forget or delete sensitive, copyrighted, or private information upon request has become crucial. Machine unlearning has been proposed to address this challenge. However, current evaluations for unlearning methods rely on output-based metrics, which cannot verify whether information is completely deleted or merely suppressed at the representation level, where suppression is insufficient for true unlearning. To address this gap, we propose a novel restoration-based analysis framework that uses Sparse Autoencoders to identify class-specific expert features in intermediate layers and applies inference-time steering to quantitatively distinguish between suppression and deletion. Applying our framework to 12 major unlearning methods in image classification tasks, we find that most methods achieve high restoration rates of unlearned information, indicating that they only suppress information at the decision-boundary level, while preserving semantic features in intermediate representations. Notably, even retraining from pretrained checkpoints shows high restoration, revealing that robust semantic features inherited from pretraining are not removed by retraining. These results demonstrate that representation-level retention poses significant risks overlooked by output-based metrics, highlighting the need for new unlearning evaluation criteria. We propose new evaluation guidelines that prioritize representation-level verification, especially for privacy-critical applications in the era of pre-trained models.

The growing concern over training data privacy has elevated the "Right to be Forgotten" into a critical requirement, thereby raising the demand for effective Machine Unlearning. However, existing unlearning approaches commonly suffer from a fundamental trade-off: aggressively erasing the influence of target data often degrades model utility on retained data, while conservative strategies leave residual target information intact. In this work, the intrinsic representation properties learned during model pretraining are analyzed. It is demonstrated that semantic class concepts are entangled at the feature-pattern level, sharing associated features while preserving concept-specific discriminative components. This entanglement fundamentally limits the effectiveness of existing unlearning paradigms. Motivated by this insight, we propose Machine-Guided Unlearning (MeGU), a novel framework that guides unlearning through concept-aware re-alignment. Specifically, Multi-modal Large Language Models (MLLMs) are leveraged to explicitly determine re-alignment directions for target samples by assigning semantically meaningful perturbing labels. To improve efficiency, inter-class conceptual similarities estimated by the MLLM are encoded into a lightweight transition matrix. Furthermore, MeGU introduces a positive-negative feature noise pair to explicitly disentangle target concept influence. During finetuning, the negative noise suppresses target-specific feature patterns, while the positive noise reinforces remaining associated features and aligns them with perturbing concepts. This coordinated design enables selective disruption of target-specific representations while preserving shared semantic structures. As a result, MeGU enables controlled and selective forgetting, effectively mitigating both under-unlearning and over-unlearning.

Machine unlearning is a promising approach to improve LLM safety by removing unwanted knowledge from the model. However, prevailing gradient-based unlearning methods suffer from issues such as high computational costs, hyperparameter instability, poor sequential unlearning capability, vulnerability to relearning attacks, low data efficiency, and lack of interpretability. While Sparse Autoencoders are well-suited to improve these aspects by enabling targeted activation-based unlearning, prior approaches underperform gradient-based methods. This work demonstrates that, contrary to these earlier findings, SAEs can significantly improve unlearning when employed dynamically. We introduce $\textbf{Dynamic DAE Guardrails}$ (DSG), a novel method for precision unlearning that leverages principled feature selection and a dynamic classifier. Our experiments show DSG substantially outperforms leading unlearning methods, achieving superior forget-utility trade-offs. DSG addresses key drawbacks of gradient-based approaches for unlearning -- offering enhanced computational efficiency and stability, robust performance in sequential unlearning, stronger resistance to relearning attacks, better data efficiency including zero-shot settings, and more interpretable unlearning.