IoT devices have become indispensable components of our lives, and the advancement of AI technologies will make them even more pervasive, increasing the vulnerability to malfunctions or cyberattacks and raising privacy concerns. Encryption can mitigate these challenges; however, most existing anomaly detection techniques decrypt the data to perform the analysis, potentially undermining the encryption protection provided during transit or storage. Homomorphic encryption schemes are promising solutions as they enable the processing and execution of operations on IoT data while still encrypted, however, these schemes offer only limited operations, which poses challenges to their practical usage. In this paper, we propose a novel privacy-preserving anomaly detection solution designed for homomorphically encrypted data generated by IoT devices that efficiently detects abnormal values without performing decryption. We have adapted the Histogram-based anomaly detection technique for TFHE scheme to address limitations related to the input size and the depth of computation by implementing vectorized support operations. These operations include addition, value placement in buckets, labeling abnormal buckets based on a threshold frequency, labeling abnormal values based on their range, and bucket labels. Evaluation results show that the solution effectively detects anomalies without requiring data decryption and achieves consistent results comparable to the mechanism operating on plain data. Also, it shows robustness and resilience against various challenges commonly encountered in IoT environments, such as noisy sensor data, adversarial attacks, communication failures, and device malfunctions. Moreover, the time and computational overheads determined for several solution configurations, despite being large, are reasonable compared to those reported in existing literature.

翻译：物联网设备已成为我们生活中不可或缺的组成部分，而人工智能技术的进步将使其更加普及，从而增加故障或网络攻击的脆弱性并引发隐私问题。加密技术可以缓解这些挑战，然而现有的大多数异常检测技术需要对数据进行解密以执行分析，这可能会削弱在传输或存储过程中提供的加密保护。同态加密方案是有前景的解决方案，因为它们能在数据仍处于加密状态时对物联网数据进行处理并执行操作，但这些方案仅支持有限的操作，这对其实际应用提出了挑战。在本文中，我们提出了一种新颖的隐私保护异常检测解决方案，专为物联网设备生成的同态加密数据设计，能够在不执行解密的情况下高效检测异常值。我们针对TFHE方案改进了基于直方图的异常检测技术，通过实现向量化支持操作来解决输入大小和计算深度相关的限制。这些操作包括加法、将值放入桶中、基于阈值频率标记异常桶、基于值范围标记异常值以及桶标签。评估结果表明，该解决方案无需数据解密即可有效检测异常，并且取得了与基于明文数据运行的机制一致的结果。此外，它对于物联网环境中常见的各种挑战（如噪声传感器数据、对抗性攻击、通信故障和设备故障）表现出鲁棒性和韧性。并且，尽管为多种解决方案配置确定的时间和计算开销较大，但与现有文献中报告的结果相比，这些开销是合理的。