Zhipeng Zhou

Yulin Zhang, He Cao, Zihao Jiang et al.

Designing proteins with desired functions or properties represents a core goal in synthetic biology and drug discovery. Recent advances in protein language models (PLMs) have enabled the generation of highly designable protein sequences, while preference alignment provides a promising way to steer designs toward desired functions and properties. Nevertheless, they often trigger catastrophic forgetting of pretrained knowledge, degrading basic designability and failing to balance multiple competing objectives. To address these issues, we draw inspiration from On-Policy Distillation (OPD), an advanced post-training method renowned for mitigating catastrophic forgetting through its mode-seeking nature. In this work, we propose ProteinOPD, a multi-objective preference alignment framework that can effectively balance multiple preference objectives while maintaining the inherent designability of PLMs. ProteinOPD adapts a pretrained PLM into preference-specific teachers and distills their knowledge into a shared student via token-level OPD on the student's own trajectories. During this process, the student is aligned to a unique normalized geometric consensus of weighted teachers while ensuring bounded optimization under conflicts. This bridges the gap for OPD in multi-objective/teacher alignment. Extensive experiments show that ProteinOPD achieves substantial gains on target preference objectives without compromising the designability, with an 8x training speedup over RL-based alignment competitors.

Zhengqiang Ge, Zhipeng Zhou, Dong Guo et al.

Neural networks, with the capability to provide efficient predictive models, have been widely used in medical, financial, and other fields, bringing great convenience to our lives. However, the high accuracy of the model requires a large amount of data from multiple parties, raising public concerns about privacy. Privacy-preserving neural network based on multi-party computation is one of the current methods used to provide model training and inference under the premise of solving data privacy. In this study, we propose a new two-party privacy-preserving neural network training and inference framework in which privacy data is distributed to two non-colluding servers. We construct a preprocessing protocol for mask generation, support and realize secret sharing comparison on 2PC, and propose a new method to further reduce the communication rounds. Based on the comparison protocol, we construct building blocks such as division and exponential, and realize the process of training and inference that no longer needs to convert between different types of secret sharings and is entirely based on arithmetic secret sharing. Compared with the previous works, our work obtains higher accuracy, which is very close to that of plaintext training. While the accuracy has been improved, the runtime is reduced, considering the online phase, our work is 5x faster than SecureML, 4.32-5.75x faster than SecureNN, and is very close to the current optimal 3PC implementation, FALCON. For secure inference, as far as known knowledge is concerned, we should be the current optimal 2PC implementation, which is 4-358x faster than other works.