Computation offloading (often to external computing resources over a network) has become a necessity for modern applications. At the same time, the proliferation of machine learning techniques has empowered malicious actors to use such techniques in order to breach the privacy of the execution process for offloaded computations. This can enable malicious actors to identify offloaded computations and infer their nature based on computation characteristics that they may have access to even if they do not have direct access to the computation code. In this paper, we first demonstrate that even non-sophisticated machine learning algorithms can accurately identify offloaded computations. We then explore the design space of anonymizing offloaded computations through the realization of a framework, called Camouflage. Camouflage features practical mechanisms to conceal characteristics related to the execution of computations, which can be used by malicious actors to identify computations and orchestrate further attacks based on identified computations. Our evaluation demonstrated that Camouflage can impede the ability of malicious actors to identify executed computations by up to 60%, while incurring modest overheads for the anonymization of computations.

翻译：计算卸载（通常通过网络将计算任务外包至外部计算资源）已成为现代应用的必然需求。与此同时，机器学习技术的普及使恶意行为者能够利用此类技术破坏卸载计算执行过程的隐私性。即便无法直接访问计算代码，恶意行为者仍可能通过获取的计算特性识别卸载计算并推测其性质。本文首先证明，即使是非高级的机器学习算法也能准确识别卸载计算。随后，我们通过实现名为Camouflage的框架探索匿名化卸载计算的设计空间。Camouflage具备实用机制，可隐藏与计算执行相关的特性——这些特性本可能被恶意行为者用于识别计算并基于识别结果策划进一步攻击。评估表明，Camouflage可将恶意行为者识别已执行计算的能力降低多达60%，同时仅引入适度的匿名化开销。