EASE: Federated Multimodal Unlearning via Entanglement-Aware Anchor Closure

The Problem: AI That Cannot Forget

Imagine a multimodal AI system trained on data from multiple hospitals, learning to associate medical images with clinical text. One day, a hospital demands: "Remove the influence of all our patient data from your model."

This is hard because:

• The model has already been trained on millions of data points
• The data is cross-modal — image and text representations are entangled in a joint embedding space
• It is a federated learning setting — data was never centralized; each client only holds its own data

This is the core challenge of federated multimodal unlearning.

Why Multimodal Unlearning Is So Difficult

Three levels of entanglement make unlearning extremely difficult:

1. Cross-modal entanglement — Deleting the influence of images may accidentally remove information from associated text, and vice versa. 2. Client gradient subspace entanglement — In federated learning, different clients' gradient updates are interleaved in parameter space. You cannot simply "subtract" one client's gradients, because its gradient directions may be shared with other clients. 3. Blurred retain/forget boundary — Coarse unlearning can damage model performance on retained data.

It is like trying to precisely remove one ingredient's flavor from an already-cooked soup.

EASE: The Anchor Principle and Entanglement-Aware Closure

The paper proposes EASE, with three key ideas:

1. Anchor Principle — Identify "unlearning-aligned" anchors: the cross-modal associations learned from the data to be forgotten. These anchors are the connections to sever — but only the unlearning-relevant ones, preserving shared ones. 2. Entanglement-aware anchor closure — Beyond severing anchors, EASE closes the cross-modal reconstruction channels they maintain, preventing forgotten information from indirectly recovering through bilinear coupling. 3. Separated update directions — Distinguish "unlearning-specific" gradient directions from those shared with retained clients, and unlearn only the former.

This acts as a precision "digital scalpel" — not cutting out an entire organ, but precisely severing specific neural connections.

Why It Matters

Demand for unlearning in multimodal federated learning is growing:

• Privacy regulations (e.g., GDPR): users have the right to request deletion of their data
• License changes: data providers may revoke usage authorization
• Erroneous data correction: training data may turn out to be wrong
• Bias removal: some data may introduce bias that needs to be "unlearned"

Takeaways

If you are building federated or multimodal AI systems, ask:

1. Does my system support precisely unlearning the influence of specific data? 2. Has cross-modal information become entangled beyond separation? 3. In a federated setting, how can unlearning be done without exposing privacy? 4. Does the unlearning operation damage performance on retained data?

EASE's philosophy: unlearning is not the opposite of learning, but a necessary complement. Learning makes a system capable; forgetting keeps it correct. In the federated multimodal world, EASE offers an elegant solution — not brute-force deletion, but precise "disentanglement."

---

Paper: EASE: Federated Multimodal Unlearning via Entanglement-Aware Anchor Closure Authors: Zihao Ding, Beining Wu, Jun Huang arXiv: 2605.00733 | 2026-04-30