Trustworthy Distributed Brain-inspired Systems: Theoretical Basis and Hardware Implementation

人工神经网络(ANN)是越来越多与关键系统集成并使用敏感用户数据的应用程序的核心，这使得安全和隐私问题变得至关重要。在机器人应用程序中损害用于对象检测的分类器可能会导致安全故障，而医疗记录等敏感数据的泄露会给用户带来隐私问题，并给提供商带来法律风险。最近，联邦学习(FL)作为一种很有前途的分布式学习方法出现，它允许从属于多个参与者的数据进行学习，而不会损害隐私，因为用户数据从不直接交换。虽然FL已被推广为一种隐私保护方法，但最近的研究表明，这种方法容易受到复杂的攻击，这些攻击可能会危及这些系统的完整性和隐私，或者以其他方式扰乱它们的操作。现有的防御系统不能覆盖FL系统面临的威胁范围，在某些情况下，防御一类攻击会增加对其他攻击的脆弱性。此外，最先进的防御需要高功率开销，这可能不适用于FL系统中的嵌入式系统和边缘节点。虽然人工神经网络是机器学习(ML)的事实上的体系结构，但像尖峰神经网络(SNNS)这样的神经形态体系结构最近已经成为一种有吸引力的替代方案，因为它们具有生物学上的可能性和大脑启发的功能。此外，神经形态硬件可以利用异步神经元的行为来实现显著的高能量效率。此外，我们的初步研究表明，与ANN相比，这些体系结构在安全性方面具有很好的优势。我们相信这些优势使得神经形态体系结构成为安全和保护隐私的低功耗分布式智能系统的潜在候选者。在TruBrain中，我们提出了对隐私保护、安全和低功耗分布式智能系统的研究努力。我们的研究目标如下：-目标1：调查神经形态节点的安全和隐私威胁，并表征其固有的安全和隐私保护特征-目标2：构建安全的大脑启发FL架构：我们利用大脑启发架构开发可证明安全的实用神经形态FL系统。-目标3：通过硬件感知的理论研究，在分布式神经形态学习系统的安全性方面弥合理论和实践之间的差距。目标4：在现场可编程门阵列上设计并实现了神经形态FL节点的硬件平台，并将其集成到RISC-V架构中。-目标5：在一个医疗应用用例中演示我们的神经形态FL范例，并从安全和隐私的角度验证其可信性。

Artificial Neural Networks (ANNs) are at the core of an increasing number of applications integrated with critical systems and using sensitive user data, making security and privacy concerns critical. Compromising a classifier for object detection in a robotic application can lead to safety breakdowns, while leakage of sensitive data, such as medical records, raises privacy concerns for users and legal exposure for providers. Recently, Federated Learning (FL) emerged as a promising distributed learning approach that enables learning from data belonging to multiple participants, without compromising privacy since user data is never directly exchanged. While FL has been promoted as a privacy-preserving approach, recent studies show that this approach is vulnerable to sophisticated attacks that are able to jeopardise both integrity and privacy of these systems, or otherwise disrupt their operation. Existing defences fall short of covering the range of threats that face FL systems, and in some cases defending against a class of attacks increases the vulnerability to other attacks. Moreover, state-of-art defences require high power overhead that might not be practical for embedded systems and Edge nodes in a FL system. While ANNs are the de-facto architectures for Machine Learning (ML), neuromorphic architectures like Spiking Neural Networks (SNNs) have recently emerged as an attractive alternative, due to their biological plausibility and brain-inspired functionality. Moreover, neuromorphic hardware can exploit the asynchronous neurons' behaviour to achieve significantly high energy efficiency. Besides, our preliminary studies show promising superiority of these architectures compared to ANNs in terms of security. We believe these advantages make neuromorphic architectures a promising candidate for secure and privacy-preserving low power distributed intelligent systems.In TruBrain, we propose a research effort towards privacy-preserving, secure and low power distributed intelligent systems. Our research objectives are as follows:- Objective1: Investigating the security and privacy threats for Neuromorphic nodes and characterising their inherent security and privacy-preserving characteristics- Objective2: Building a secure brain-inspired FL architecture: We leverage brain-inspired architectures to develop provably-secure practical neuromorphic FL systems.- Objective3: Bridging the gap between theory and practice in distributed neuromorphic learning systems' security through a hardware-aware theoretical study. - Objective4: Designing and implementing a Hardware platform for neuromorphic FL nodes on FPGA and integrating it in a RISC-V architecture. - Objective 5: Demonstrating our neuromorphic FL paradigm in a medical application use case, and validating its trustworthiness from a security and privacy perspective.