Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks

2026-07-08 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, quick

Summary

A new survey provides a unified, system-oriented view of multimodal unlearning across vision, language, audio, and video. This field addresses the challenge of foundation models like VLMs, DMs, LLMs, and AFMs inadvertently encoding sensitive, copyrighted, biased, or unsafe cross-modal associations from their training data. Retraining these models after deletion requests or policy updates is often impractical, and targeted forgetting is difficult due to knowledge distribution across shared representations. Multimodal unlearning enables selective removal of specific information across modalities while retaining overall model utility. The survey establishes a taxonomy for systematic comparison across model architectures and modalities, clarifying trade-offs among deletion strength, retention, efficiency, reversibility, and robustness. It also highlights open problems and practical considerations to guide future research and deployment. A curated repository is available at https://smsnobin77.github.io/Awesome-Multimodal-Unlearning/.

Key takeaway

For AI Scientists and Machine Learning Engineers developing or deploying multimodal foundation models, you must consider robust multimodal unlearning strategies. This is crucial for addressing inadvertent encoding of sensitive or biased data without costly full model retraining. You should evaluate unlearning methods based on the survey's taxonomy, prioritizing deletion strength, efficiency, and retention to ensure compliance and model integrity. Explore the curated repository to inform your implementation decisions.

Key insights

Multimodal unlearning selectively removes sensitive data from foundation models while preserving utility.

Principles

• Knowledge in foundation models is distributed across shared representations.
• Retraining for data deletion is often impractical.
• Unlearning involves trade-offs: strength, retention, efficiency, reversibility, robustness.

Method

The survey proposes a taxonomy for comparing multimodal unlearning methods based on model architectures and modalities, evaluating deletion strength, retention, efficiency, reversibility, and robustness.

In practice

• Address sensitive data encoding in VLMs, DMs, LLMs, AFMs.
• Implement selective data removal for policy compliance.
• Evaluate unlearning methods using the proposed taxonomy.

Topics

• Multimodal Unlearning
• Foundation Models
• Data Privacy
• Model Debiasing
• Machine Unlearning
• VLMs, LLMs, DMs, AFMs

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer

Related on AIssential

• Multimodal Unlearning Across Vision, Language, Video, and Audio: Survey of Methods, Datasets, and Benchmarks — Multimodal unlearning selectively removes unwanted data associations from foundation models without costly retraining.
• Auditing Machine Unlearning: A Systematic Research on Whether Models Truly Forget — A new auditing framework reliably verifies machine unlearning effectiveness, revealing varying success across different unlearning methods.
• POPS: Recovering Unlearned Multi-Modality Knowledge in MLLMs with Prompt-Optimized Parameter Shaking — POPS demonstrates that current Multi-modality Machine Unlearning methods are fundamentally vulnerable to adversarial recovery of sensitive data.
• Unlearning Under Imbalance: Benchmarking Fairness in Multimodal LLM Unlearning — Unbalanced unlearning requests in MLLMs can induce bias, necessitating benchmarks like FAIRGET and algorithms like FAUN to ensure fairness.
• Stochastic Meta-Unlearning: Bridging Language Backbone and Multimodal Unlearning — VLM unlearning requires multimodal feedback; text-only unlearning on the language backbone is insufficient for complete forgetting.
• Fair Cognitive Impairment Detection Through Unlearning — Demographic unlearning via gradient reversal in multimodal fusion improves fair and robust MCI detection from speech.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.