Recent work has highlighted the risks of intellectual property (IP) piracy of deep learning (DL) models from the side-channel leakage of DL hardware accelerators. In response, to provide side-channel leakage resiliency to DL hardware accelerators, several approaches have been proposed, mainly borrowed from the methodologies devised for cryptographic implementations. Therefore, as expected, the same challenges posed by the complex design of such countermeasures should be dealt with. This is despite the fact that fundamental cryptographic approaches, specifically secure and private function evaluation, could potentially improve the robustness against side-channel leakage. To examine this and weigh the costs and benefits, we introduce hardware garbled NN (HWGN2), a DL hardware accelerator implemented on FPGA. HWGN2 also provides NN designers with the flexibility to protect their IP in real-time applications, where hardware resources are heavily constrained, through a hardware-communication cost trade-off. Concretely, we apply garbled circuits, implemented using a MIPS architecture that achieves up to 62.5x fewer logical and 66x less memory utilization than the state-of-the-art approaches at the price of communication overhead. Further, the side-channel resiliency of HWGN2 is demonstrated by employing the test vector leakage assessment (TVLA) test against both power and electromagnetic side-channels. This is in addition to the inherent feature of HWGN2: it ensures the privacy of users' input, including the architecture of NNs. We also demonstrate a natural extension to the malicious security modeljust as a by-product of our implementation.
翻译:最近的工作突出了知识产权(IP)盗用DL硬件加速器侧通道渗漏产生的深层学习(DL)模型的风险。作为回应,为向DL硬件加速器提供侧通道渗漏复原力,提出了几种办法,主要是从为加密实施设计的方法中借用,因此,正如预期的那样,应当处理这类对策的复杂设计所构成的同样挑战。尽管基本的加密方法,特别是安全和私用功能评估,有可能提高防侧通道渗漏的稳健性能。为了检查这一点并权衡成本和效益,我们为DL硬件加速器提供了侧通道渗漏耐DL硬件加速器(HWGN2),一个DL硬件加速器在FGA上实施。HWG2还使NN设计者在实时应用中保护其IP方面拥有灵活性,而硬件资源由于硬件通信成本交易而受到严重制约。具体地说,我们采用混合电路路,使用MIPS结构实现最高至62.5x水平的内层漏漏漏漏漏漏漏损(HWG的逻辑和66x内层内层内层内层内层内层内层内层内层内层内层内层内层内层内压系统内压,也通过进一步测试系统内压系统内压,使HWG的内层内层内压系统内压系统内压系统内压系统内压系统内压系统内置。)的内压。