Machine Unlearning
Government of India (SERB) sponsored initiative (₹30.5 Lakh) investigating algorithmic data removal and privacy compliance in deep learning models without performance degradation. Successfully completed on January 1, 2026.
Science and Engineering Research Board (SERB)
Statutory Body Established Through an Act of Parliament: SERB Act 2008 • Department of Science & Technology (DST), Government of India
Regulatory Motivation & Problem Statement
In contemporary deep learning systems, neural networks memorize intricate patterns directly from training distributions. However, emergent regulatory frameworks—including the European Union's General Data Protection Regulation (GDPR Article 17: "Right to be Forgotten"), the California Consumer Privacy Act (CCPA), and India's Digital Personal Data Protection (DPDP) Act 2023—legally mandate that individuals can revoke consent and demand complete removal of their personal data footprints from trained predictive models.
The standard naive solution—retraining massive models from scratch while excluding the targeted data shards—is computationally prohibitive, financially unsustainable, and environmentally costly when scaling across billions of parameters. This research project, funded under the prestigious SERB Startup Research Grant (SRG), investigated principled algorithmic frameworks to achieve exact and approximate machine unlearning. The objective was to surgically excise specific data influence from deep neural representations rapidly, while strictly preserving task utility and preventing catastrophic forgetting on retained distributions.
Core Research Objectives
The project established both theoretical bounds and scalable empirical algorithms for efficient data erasure across dense deep learning architectures:
Targeted Parameter Scrubbing
Designing efficient first- and second-order weight adjustment algorithms that nullify the influence of requested samples without full model retraining.
Catastrophic Forgetting Prevention
Formulating regularized feature constraints to guarantee that the decision boundaries of retained classes maintain high fidelity and zero-shot generalization.
Empirical Auditing & Verification
Developing mathematical auditing protocols based on Membership Inference Attacks (MIA) to empirically guarantee that deleted data cannot be reconstructed.
Architecture Scalability
Validating algorithmic efficiency across modern backbones including ResNets, Vision Transformers (ViTs), and multi-modal representation pipelines.
Phased Milestones & Project Execution
With the full sanction and 100% disbursement of ₹30,50,300 INR, all project objectives and milestones were successfully executed and concluded on January 1, 2026:
Theoretical Framing & Baseline Verification (Jan 2024 – Jun 2024)
Mathematical formalization of unlearning criteria; deployment of standard computer vision benchmark datasets; establishment of baseline Membership Inference auditing pipelines.
Algorithm Development & Parameter Scrubbing (Jul 2024 – Jun 2025)
Implementation of selective gradient inversion and localized Hessian approximations; extensive ablation on memory footprints and compute runtime reductions compared to standard retraining.
Large-Scale Validation, Auditing Suite & Dissemination (Jul 2025 – Jan 2026)
Benchmarking across vision-language architectures; development of empirical unlearning verification methodologies; publication dissemination and final SERB grant report submission.
Research Personnel & Laboratory
This project was conducted under the Responsible and Trustworthy Artificial Intelligence Lab at the School of Computer Applications, KIIT Deemed to be University:
Dr. Murari Mandal
Umakanta Maharana
Foundational Literature & References
Key theoretical foundations and seminal literature informing the algorithmic development of this project:
Academic Collaborations & Inquiries
We welcome inquiries from academic researchers, students, and industry partners interested in machine unlearning, data privacy, and trustworthy AI.