Adversarial-HD: Hyperdimensional Computing Adversarial Attack Design for Secure Industrial Internet of Things
Adversarial-HD: Hyperdimensional Computing Adversarial Attack Design for Secure Industrial Internet of Things
复制标题
DOI：
10.1145/3576914.3587484
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Onat Gungor;T. Rosing;Baris Aksanli
中科院分区：
文献类型：
--
作者：
Onat Gungor;T. Rosing;Baris Aksanli
工业物联网 (IIoT) 是传感器、网络设备和从工业运营中收集数据的设备的协作。由于互连性和有限的计算能力，工业物联网系统具有许多安全漏洞。基于机器学习的入侵检测系统（IDS）是一种可能的安全方法，它可以持续监控网络数据并以自动方式检测网络攻击。超维 (HD) 计算是一种受大脑启发的机器学习方法，它足够准确，同时极其稳健、快速且节能。基于这些特性，HD 可以成为适用于 IIoT 系统的基于 ML 的 IDS 解决方案。然而，其预测性能会受到输入数据中的小扰动的影响。为了充分评估HD的漏洞，我们提出了一种有效的面向HD的对抗攻击设计。我们首先选择最多样化的攻击集，以最大限度地减少开销并消除对抗性冗余。然后，我们执行实时攻击选择，找出最有效的攻击。我们在真实的 IIoT 入侵数据集上进行的实验表明了我们的攻击设计的有效性。与最有效的单一攻击相比，我们的设计策略可以将攻击成功率提高高达36%，F1分数提高高达61%。
Industrial Internet of Things (IIoT) is a collaboration of sensors, networking equipment, and devices to collect data from industrial operations. IIoT systems possess numerous security vulnerabilities due to inter-connectivity and limited computational power. Machine learning based intrusion detection system (IDS) is one possible security approach that continuously monitors network data and detects cyberattacks in an automated manner. Hyper-dimensional (HD) computing is a brain-inspired ML method that is sufficiently accurate while being extremely robust, fast, and energy-efficient. Based on these characteristics, HD can be a favorable ML-based IDS solution for IIoT systems. However, its prediction performance is impacted by small perturbations in the input data. To fully evaluate the vulnerabilities of HD, we propose an effective HD-oriented adversarial attack design. We first select the most diverse set of attacks to minimize overhead, and eliminate adversarial redundancy. Then, we perform a real-time attack selection which finds out the most effective attack. Our experiments on a realistic IIoT intrusion data set show the effectiveness of our attack design. Compared to the most effective single attack, our design strategy can improve attack success rate by up to 36%, and F1 score by up to 61%.