New Framework Lets AI Systems Selectively Delete Sensitive Data Modalities On Demand

As multimodal AI systems collect and process increasingly sensitive information—video, audio, images, text—controlling what data they retain has emerged as a key privacy challenge. A new research paper introduces Missing-by-Design (MBD), a framework that allows users or regulators to revoke access to specific data modalities from trained sentiment analysis systems without retraining the entire model.

The Privacy Problem at Scale#

Multimodal AI systems that analyze sentiment often rely on multiple input types: images, audio, text, or video. When users request deletion of their data—or when regulations like GDPR mandate removal—organizations typically face a stark choice: keep the model intact and violate privacy rules, or retrain from scratch, which is costly and time-consuming. MBD proposes a middle path: surgical unlearning that removes just the unwanted modality.

How Missing-by-Design Works#

The framework combines two core mechanisms. First, it learns property-aware embeddings—internal representations designed to separate signals relevant to the task from information tied to specific modalities. When data must be deleted, a generator attempts to reconstruct missing channels from remaining ones, allowing the model to continue functioning with partial inputs.

Second, for formal deletion requests, MBD employs a saliency-driven selection and calibrated Gaussian update. This produces a Modality Deletion Certificate—a machine-verifiable proof that a given modality's influence has been removed from the model. The technical process identifies the most important parameters tied to a specific data type, then modifies them in a controlled way that the system can cryptographically verify.

Performance Trade-offs#

Experiments on benchmark datasets demonstrate that the approach delivers competitive predictive accuracy even with incomplete inputs, while achieving a practical balance between privacy protection and model utility. Rather than full retraining, the method positions unlearning as an efficient, verifiable alternative—reducing computational cost and enabling rapid compliance with deletion requests.

The work addresses a growing regulatory and ethical demand: giving users genuine control over their data in systems that have already learned from it.