Quantized Neural Networks (QNNs) have received increasing attention, since they can enrich intelligent applications deployed on embedded devices with limited resources, such as mobile devices and AIoT systems. Unfortunately, the numerical and computational discrepancies between training systems (i.e., servers) and deployment systems (e.g., embedded ends) may lead to large accuracy drop for QNNs in real deployments. We propose a Computation-Quantized Training Framework (CQTF), which simulates deployment-time fixed-point computation during training to enable one-shot, lossless deployment. The training procedure of CQTF is built upon a well-formulated quantization-specific numerical representation that quantifies both numerical and computational discrepancies between training and deployment. Leveraging this representation, forward propagation executes all computations in quantization mode to simulate deployment-time inference, while backward propagation identifies and mitigates gradient vanishing through an efficient floating-point gradient update scheme. Benchmark-based experiments demonstrate the efficiency of our approach, which can achieve no accuracy loss from training to deployment. Compared with existing five frameworks, the deployed accuracy of CQTF can be improved by up to 18.41%.

Bit-Trojan attacks based on Bit-Flip Attacks (BFAs) have emerged as severe threats to Deep Neural Networks (DNNs) deployed in safety-critical systems since they can inject Trojans during the model deployment stage without accessing training supply chains. Existing works are mainly devoted to improving the executability of Bit-Trojan attacks, while seriously ignoring the concerns on evasiveness. In this paper, we propose a highly Evasive Targeted Bit-Trojan (ETBT) with evasiveness improvements from three aspects, i.e., reducing the number of bit-flips (improving executability), smoothing activation distribution, and reducing accuracy fluctuation. Specifically, key neuron extraction is utilized to identify essential neurons from DNNs precisely and decouple the key neurons between different classes, thus improving the evasiveness regarding accuracy fluctuation and executability. Additionally, activation-constrained trigger generation is devised to eliminate the differences between activation distributions of Trojaned and clean models, which enhances evasiveness from the perspective of activation distribution. Ultimately, the strategy of constrained target bits search is designed to reduce bit-flip numbers, directly benefits the evasiveness of ETBT. Benchmark-based experiments are conducted to evaluate the superiority of ETBT. Compared with existing works, ETBT can significantly improve evasiveness-relevant performances with much lower computation overheads, better robustness, and generalizability. Our code is released athttps://github.com/bluefier/ETBT.