CleanSheet: Advancing backdoor attack techniques for deep neural networks with stealthy trigger embedding

Backdoor attacks pose a significant threat to the security of deep neural networks by enabling hidden manipulations that alter model predictions when specific triggers are present. Many existing attacks struggle with limited transferability across architectures, reduced stealth, or vulnerability to detection by current defense methods. This work introduces CleanSheet, a novel backdoor attack framework that addresses these challenges through compact, adaptive trigger designs. CleanSheet is evaluated across a wide range of datasets CIFAR-10, AG News, SVHN, TinyImageNet, IMDB and MalNet-Tiny and models including ResNet-18, VGG-16, DenseNet-121, BERT, and GPT-3. It achieves an attack success rate of up to 96.2% on GPT-3 with the IMDB dataset while maintaining high accuracy on clean inputs. CleanSheet also bypasses advanced defenses such as ONION, input sanitization, and anomaly detection, consistently achieving over 89% attack success even under defense. We further analyze how factors like trigger size and type, dataset scale, training duration, and model complexity affect the attack’s performance. Compared to baseline methods, CleanSheet improves attack success rate by an average of 12.4%. These results highlight the effectiveness and stealth of CleanSheet, calling attention to the need for improved defense mechanisms in machine learning systems.