Exploring the Application of Diffusion Model Algorithms in the EDA Chip Field. To improve the efficiency and quality of fault diagnosis, innovatively transform test feedback into images, design multiple machine learning algorithms combined with predictive dynamic termination of testing, and use the efficient diffusion model Denoising Diffusion Implicit Model (DDIM) to expand the image dataset for training.

• Implement DDPM, DDIM, and D3PM models on the existing chip circuit fault log-picture dataset, sample fault data pictures to supplement the original picture dataset, forming a new synthetic dataset.
• Train the yolo8n-cls model for test termination prediction classification tasks on both the original and synthetic datasets, compare and evaluate the two models, and verify the data augmentation effect. The results show that the new scheme has a higher accuracy rate than the original scheme, with an expected test cost saving of nearly 90%.
• The team’s paper “Translating Test Responses to Images for Test-termination Prediction via Multiple Machine Learning Strategies” submitted to ACM-TODEAS (CCF-B).

Understand basic parallel programming methods, program and optimize the performance of shared memory and distributed memory parallel computer systems. Understand CUDA programming optimization designed for GPUs.

• Master the basic methods of parallel optimization algorithm design, and implement computer performance optimization based on methods such as OpenMP and MPI. Compared with traditional methods, achieve 2 to 4 times performance improvement in key issues such as matrix operations.

Research on medical image segmentation models. In response to the complex model structure, high computational cost, and difficulty in direct deployment on resource-constrained medical device terminals, explore and integrate Quantization-Aware Training (QAT) for model compression and inference acceleration, quantize the segmentation model to 8-bit, and reduce the computational volume to about 1/4 of the floating-point precision, significantly reducing the computational volume with almost no loss of accuracy.

• Explore the application of federated learning in multi-institutional tumor segmentation systems, integrate PET-CT dual-modality images for learning, introduce the self-attention LSA module and adversarial noise perturbation ANP module to optimize the training of federated learning models, and achieve a significant improvement compared to the baseline model.
• To address the issue of limited resources and computing power on edge devices, use Quantization-Aware Training to compress the model. Adopt asymmetric uniform quantization for weights, and the precision loss is smaller compared to the Post-Training Quantization (PTQ) method.
• The paper in preparation is expected to be submitted to IEEE-BIBM (CCF-B).

Design and implement an intelligent detection system for multi-element audio attacks facing voice recognition systems, which has good performance in detecting synthetic attacks, replay attacks, and adversarial attacks, is user-friendly, and supports multi-end access.

• In response to the vulnerability and complexity of the environment in voice spoofing attack recognition, develop an efficient voiceprint feature extraction and analysis system based on the residual network. Integrate the new residual network Res2Net with the squeeze-and-excitation SE module to capture key voice features in detail and deeply analyze the global characteristics of voice signals. Trained and tested on the international ASVSpoof dataset, it has higher accuracy and generalization than the baseline model in recognizing synthetic attacks and replay attacks, with an EER less than 1%.
• Use the generative adversarial network to construct a neural voice coder to reconstruct voice waveforms, and use the difference in ASV matching scores between the reconstructed audio and the original audio to identify adversarial samples of unknown attack methods. Tested on the VoxCeleb open source dataset, the AUC reaches 99%, with high recognition accuracy.
• The work have been submitted to the National College Students’ Information Security Competition.