Security of Neural Networks
Defending deep models against adversarial examples, backdoor and data-poisoning attacks — and understanding when and why they fail.
Research studies, software and prototypes from across the NPerseus team.
Research studies, prototypes and software described in public profiles. Status, roles and funding follow each source record; a profile entry does not by itself establish a released product.
74 matching profile project records · Page 1 of 7
AI-Assisted Multimodal Framework for Earthquake Vulnerability Assessment of Structures — Co-PI, National Institute of Disaster Management, ₹7.986 lakh (Ongoing; PI: Shilpa Pal)
Development of AI-Based Solution for Analysing, Benchmarking and Quality Monitoring of Cybersecurity Audit Reports and Performance Monitoring of Auditing Organisations — Co-PI, MeitY, ₹3 crore (Ongoing; PI: Pawan Singh Mehra)
Development of Deep Learning based Multi-Sensor Fusion Framework for Multi-Target Surveillance System — PI, DTU Young Faculty Grant, ₹5 lakh (Ongoing)
Development of immersive content forensic marking and watermarking technology; PI Ki-Hyun Jung; Rajeev Kumar, researcher; Korean Research Foundation; 24 million KRW; completed per current DTU profile
Intelligent Surgical Scheduling and Doctor Roster Synchronisation System with Automated Patient Reminders for Multi-Speciality Eye Care Facilities/Hospitals — Principal Investigator; funded by the DTU Centre for Community Development and Research (DTU/CCDR), in collaboration with Dr. Shroff’s Charity Eye Hospital, Delhi.
RECAP — Reversible Embedding with Causal Assistance and Prediction; PI; DTU Young Faculty Grant; ₹5 lakh; approved 21 March 2025; currently listed ongoing
SMIF-Net: A Robust Multi-Domain Learning Framework for Social Media Image Forensics — Co-Principal Investigator; Principal Investigator: Dr. Satyabrata Adhikari; Faculty Interdisciplinary Research Project (FIRP), DTU; project award approved by office order DTU/R&D/Faculty Proposal/2026/3632 dated 1 September 2026.
TagExDF — Explained Tagging System; Ki-Hyun Jung and Rajeev Kumar listed as PI/Co-PI; NRF Korea; 55 million KRW; sanctioned November 2021; completed (status confirmed by the lab on 5 September 2026).
co-developed a 3D convolutional autoencoder combining residual learning and hybrid channel attention to remove mixed noise from hyperspectral images.
developed a network for robust classification using multi-source features.
studies model behavior and gradient-entropy signals for detecting backdoored models from adversarial inputs.
Defending deep models against adversarial examples, backdoor and data-poisoning attacks — and understanding when and why they fail.
Reversible data hiding in encrypted images, image and text steganography, and robust watermarking for ownership and integrity.
Detecting manipulated and GAN-generated media — deepfakes, splicing and median-filtering traces.
Federated and privacy-preserving learning, explainable and trustworthy models, and methods for limiting the disclosure of training data.