AI-powered autonomous driving systems are being increasingly adopted by automobile manufacturers, while training autonomous driving models using machine learning technique requires vast amounts of driving data. Manufacturers tend to collaborate through ecosystem partnerships to jointly advance the development of autonomous driving algorithms such as Huawei has established partnerships with multiple collaborators sharing intelligent automotive solutions. However, concerns about data privacy naturally arise among automobile manufacturers. This paper proposes a privacy-preserving framework, EPAuto, for collaborative machine learning among automobile manufacturers. EPAuto leverages secure multi-party computation (MPC) to enable efficient privacy-preserving machine learning (PPML). We propose our improved bit-wise comparison and most significant bit (MSB) extraction protocols, leveraging the random permutation matrix as the core technique. Additionally, we design fundamental building blocks for privacy-preserving machine learning (PPML), including DReLU, ReLU, and Maxpool protocols. The security of EPAuto is formally proven under the static semi-honest adversary model in the dishonest-majority setting. We implement our protocols within 2 to 120 parties in both LAN and WAN environments. Our results demonstrate nanosecond-level online response over LAN in multi-party settings, and our PPML implementations achieve up to a 907× performance improvement compared to state-of-the-art works in the same setting.
In recent years, Federal Learning has received much attention becourse it can train models by updating gradients without contacting users’ true data. However, adversaries also can track users’ privacy from the shared gradient. In this paper, we aim to solve three major issues during the process of federated learning: 1) how to protect users’ privacy during training; 2) How to verify the correctness of the aggregation results returned from the server; 3) How to reduce communication costs while ensuring training security. So we propose a verifiable aggregation scheme that can effectively verify the results of server aggregation. Specifically, we follow the classic double mask aggregation scheme, and use Paillier homomorphic encryption algorithm to implement the message authentication code with additive homomorphic property. Users can compare their local codes with server’s aggregation results to verify the correctness of the aggregation results and improve model’s accuracy. In our framework, we adopt a Top-k gradient selection scheme to reduce models’ communication and computing overhead. Experimental results indicate that our training framework is feasible and efficient.
With the wide application of Internet of Things, big data and other technologies in smart agriculture, more and more agricultural data are collected and used. The current analysis and processing through IoT platforms and other centralised methods have more serious data security problems. Privacy computing is a technology that completes data security calculations under the premise that both parties do not disclose data, and the data of each participant participates in the calculation is in the form of cipher text to maximally reduce the risk of data leakage. Aiming at the small volume and high frequency scenario of agricultural data, this paper proposes a distributed privacy computing framework by integrating cryptography, edge computing and blockchain, and designs a privacy computing system called PCS-ADS for smart agricultural data. This paper focuses on describing the system model architecture and system function implementation, which includes key business processes, core modules, core algorithms and other parts. Finally, the security and performance evaluation of the system is carried out. The results show that, for sum and variance, the computation time for multi-party under the amount of 1w data of each party is no more than 0.1s, and for the high-frequency computation scenarios, the computation time under the amount of 100 data of each party is no more than 7s and 1.5s, respectively. Experimental results prove that PCS-ADS is effective and has a strong practicality.
Since the popularity of Machine Learning as a Service (MLaaS) has been increasing significantly, users are facing the risk of exposing sensitive information that is not task-related. The reason is that the data uploaded by users may include some information that is not useful for inference but can lead to privacy leakage. One straightforward approach to mitigate this issue is to filter out task-independent information to protect user privacy. However, this method is feasible for structured data with naturally independent entries, but it is challenging for unstructured data. Therefore, we propose a novel framework, which employs a spectrum-based encoder to transform unstructured data into the latent space and a task-specific model to identify the essential information for the target task. Our system has been comprehensively evaluated on three benchmark visual datasets and compared to previous works. The results demonstrate that our framework offers superior protection for task-independent information and maintains the usefulness of task-related information.