Can Vision Models Truly Forget? Mirage: Representation-Level Certification of Visual Unlearning
arXiv:2605.20282v1 Announce Type: cross Abstract: Machine unlearning in Vertical Federated Learning (VFL) has attracted growing interest, yet existing methods certify forgetting solely using output-level metrics. We challenge these claims by introducing Mirage, a representation-level auditing framework comprising four complementary diagnostics: Linear Probe Recovery (LPR), Centered Kernel Alignment (CKA), Feature Separability Scoring, and Layer-Wise Recovery Analysis. Through experiments across seven datasets and seven baseline methods following recent VFL unlearning protocols, Mirage reveals
The increasing focus on machine unlearning for privacy, regulatory compliance, and security in AI systems makes robust certification methods critical at this stage of AI development.
This research introduces a novel framework to rigorously audit AI model unlearning, moving beyond superficial metrics to examine representation-level changes, which is crucial for building trustworthy AI.
The ability to truly verify that sensitive information has been erased from AI models, rather than just hidden, will fundamentally change how AI systems are developed, deployed, and regulated, especially in federated learning environments.
- · AI ethicists and regulators
- · Organizations requiring data privacy compliance
- · Users of federated learning systems
- · Developers of robust unlearning algorithms
- · AI systems with superficial unlearning claims
- · Developers neglecting representation-level security
- · Organizations exposed to data leakage through unlearned models
More rigorous standards and benchmarks for machine unlearning will emerge, pushing developers towards more robust methods.
Increased trust in AI systems handling sensitive data will accelerate adoption in privacy-critical sectors like healthcare and finance.
The development of 'unlearnable' AI, where models are designed from inception to facilitate efficient and verifiable unlearning, could become a new frontier in AI architecture.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI