Bo Liu

Poisoning attacks got you down? This defense flips the script by using the attacker's own clustering behavior against them, achieving near-perfect attack mitigation with minimal accuracy loss.

GANs can now edit faces with disentangled attributes *without* needing labeled data, rivaling supervised methods in quality and speed.

Achieve state-of-the-art brain tumor segmentation even with missing MRI data by explicitly disentangling anatomical features from modality-specific biases using causality.

LLMs can be made to reason much better by directly optimizing their pre-training output distribution, even before fine-tuning on specific tasks.

You can train a surprisingly effective AI-generated image detector with just 320 images by adversarially unlearning biases towards specific generative models.

Counterintuitively, throwing more user prompts at prompt optimization can backfire, especially on diverse datasets, but a clever filtering strategy can unlock surprisingly strong generalization from just a handful of examples.

WeatherRemover achieves state-of-the-art multi-weather image restoration while drastically reducing computational cost, proving that high performance doesn't have to break the bank.

Semantic segmentation models, even recent transformer-based architectures like SAM, are surprisingly vulnerable to new backdoor attacks that current defenses can't reliably stop.