With the rapid development of large language models, the fine-tuning of downstream tasks has become a key problem. As a new learning paradigm, prompt learning can greatly reduce the number of updated parameters and effectively reduce the communication cost in federated learning scenarios by freezing the pre-trained model and fine-tuning soft prompt parameters. However, for the federated prompt tuning, the local prompt usually contains the user's personal information, and the text input of the user are easily leaked through the aggregation directly on the server. Meanwhile, the soft prompt is vulnerable to poisoning attacks launched by malicious users (such as backdoor). For the above problems, this paper proposes a privacy-preserving federated prompt fine-tuning against poisoning attacks. Specifically, we use homomorphic encryption to ensure the confidentiality of local prompts, and aggregate prompts under ciphertext to generate global prompts. Secondly, the proposed mechanism of backdoor detection is used to judge whether the prompt contains backdoors to resist the malicious attacks in the form of ciphertext. Both local and global prompts are invisible to the server during the entire training process. Finally, theoretical analysis and experimental results show that the accuracy error between the plaintext domain and the ciphertext domain is controlled within 2%-5% under backdoor attack, which effectively enhances the robustness of the model.

Outsourced inference services have greatly promoted the popularization of deep learning, and neural network models can help users customize a series of personalized applications, e.g., face recognition, image classification, etc. However, untrustworthy service providers also bring a variety of security issues, such as data privacy, network model privacy, etc. Meanwhile, the malicious example will result in the output of model incorrectly. For the above concerns, this paper proposes a privacy-reserving neural network inference against adversarial example attack (PPNNI) under two non-colluding cloud servers, where a series of efficient security protocols are designed to realize private prediction based on the additive secret sharing protocol. In the semi-honest model, the servers implement the neural network's private inference in the context of an unknown input and network model. Moreover, in order to prevent malicious examples from affecting the model accuracy, we produce a method of kernel density estimation and uncertainty value in the last layer of the hidden layer to determine whether the input is adversarial examples in the domain of ciphertext. The proposed PPNNI protocol is the first effort to efficiently detect adversarial attack behaviors under the ciphertexts. The theoretical analysis illustrate that the protocol can guarantee the privacy of input data, model parameters and the inference result. The results of the experiments demonstrate that the model accuracy are about 89% and 83% respectively with malicious examples on MNIST and CIFAR10 in the domain of ciphertext and they are nearly same as 91% and 85% in the domain of plaintext, which shows the effectiveness of our protocol.