DESIGN AND DEVELOPMENT OF UNTARGETED ATTACKS IN DNN

Abstract

Deep neural networks have become the foundation of modern artificial intelli gence systems applied to computer vision, natural language processing, autonomous sensing, medical imaging, code generation, and multimodal decision making. Their exceptional performance, however, is accompanied by increasing exposure to adver sarial threats. Among these, backdoor attacks represent a particularly serious secu rity vulnerability. A backdoored model behaves normally on clean inputs but exhibits attacker-induced erroneous behaviour when a concealed trigger is present. This dual mode operation allows malicious functionality to remain undetected during standard evaluation while posing genuine risks upon deployment. This thesis examines backdoor security in deep neural networks through two distinct but complementary angles. The first contribution investigates TWIN, a fine tuning-centred survey framework that analyses fine-tuning as both a weapon and a shield in backdoor security.Modern artificial intelligence practice increasingly relies on pre-trained models, modelhubs, plug-in adapters, prompt tuning, LoRA-style parameter efficient fine-tuning, and downstream adaptation. In this context, fine-tuning may mit igate backdoors but may also activate, preserve, or implant them. TWIN organises attacks and defences around this dual role and provides taxonomies for fine-tuning related attacks, fine-tuning-based defences, PEFT-centric vulnerabilities, adaptation stage monitoring, weight-space repair, and representation-level purification. The second contribution presents ProAttack, an advanced untargeted backdoor attack that causes semantically plausible misclassifications rather than arbitrary errors. In contrast to fixed-target attacks, ProAttack guides triggered inputs toward visually coherent wrong classes by leveraging CLIP ViT-B/32 embeddings, class prototype based destination selection, and a jointly trained U-Net perturbation generator. Bench mark evaluations across CIFAR-10, GTSRB, and CIFAR-100 confirm strong attack effectiveness with negligible clean accuracy loss and imperceptible trigger patterns. By combining these two contributions, the thesis offers a unified view of con temporary backdoor security. ProAttack demonstrates how semantic guidance can strengthen untargeted attacks, while TWIN clarifies why the fine-tuning stage has be come a critical control point in the model lifecycle. The thesis further discusses the implications of these contributions for secure deployment, defence design, trustworthy pre-trained models, and future research directions.