Junhao Wujunhao wu

Serving as a Research Intern at OV3Lab, under the supervision of Professor Jiahuan Zhou. We systematically analyzed the performance degradation of OpenVLA under visual distribution shifts and attributed it to environmental perturbations in visual representations. Based on this insight, we developed a self-supervised training framework for the vision encoder, introducing a small number of learnable prompt tokens optimized with a redundancy-reduction objective. Without requiring target-domain data or language model updates, the method effectively enhances visual invariance. Experiments on LIBERO-plus demonstrate consistent robustness improvements under lighting, layout, and background shifts, along with strong zero-shot generalization capability.

Contributed to a multimodal deepfake detection systemd that fuses semantic, frequency-domain, and spatial-domain features for social platform content moderation and evidence extraction. Worked with 100K+ data samples across multiple modalities; incorporated CLIP-based semantic features, a frequency-branch module (Freq-Net), and a spatial-branch module (CameraPrint), and applied mutual-consistency constraints and contrastive learning to improve fusion. Achieved 0.97 AUC on the validation set and 93%+ accuracy, delivering ~8 percentage-point improvement over a baseline model;