Securing Approximate Computing Systems via Obfuscating Approximate-Precise Boundary
Securing Approximate Computing Systems via Obfuscating Approximate-Precise Boundary
复制标题
DOI：
10.1109/tcad.2022.3168261
复制
发表时间：
2023-01
影响因子：
2.9
通讯作者：
Pruthvy Yellu;Qiaoyan Yu
中科院分区：
文献类型：
--
作者：
Pruthvy Yellu;Qiaoyan Yu
近似计算(AC)技术已被用来提高计算性能和能源效率，但精确度略有下降。最近的文献表明，一些AC机制可能被攻击者利用来实现新的攻击面。为了应对AC系统中新出现的攻击，我们提出了用入口模糊和边界加宽来混淆近似精确边界(APB)。该算法利用一种与近似误差有很强相关性的隐含质量度量来掩盖APB的进入，消除近似模式和精确模式之间的显式转换，从而提高交流系统对APB攻击的抵抗能力。所提出的边界加宽方案通过将单个APB阈值扩展到两个比较阈值来扩大近似模式和精确模式之间的过渡区，并进一步允许在候选库中随机选择AC模块。我们的混淆方法提供的保护机制增强了AC系统对APB攻击的弹性。我们的案例研究表明，提出的条目模糊方案比基线提高了高达168%的应用质量，并成功地达到了预期的精度。该方法的延迟开销可以忽略不计，面积和功耗的增加可以分别减少到6%和8%。
Approximate computing (AC) techniques have been leveraged to improve computing performance and energy efficiency with minor degradation in accuracy. Recent literature indicates that some AC mechanisms could be exploited by attackers to implement new attack surfaces. To address the emerging attacks in AC systems, we propose to obfuscate the approximate-precise boundary (APB) with entry-blurring and boundary-broadening schemes. The proposed entry-blurring scheme leverages a hidden quality metric, which has a strong correlation with approximation errors, to obscure the entrance of APB and eliminate the explicit transition between approximate and precise modes, thus improving AC systems’ resilience against APB attacks. The proposed boundary-broadening scheme enlarges the transition zone between approximate and precise modes by expanding a single APB threshold to two comparison thresholds, and it further enables a random selection of AC modules in the candidate library. The protection mechanisms provided by our obfuscation method strengthens AC systems’ resilience against APB attacks. Our case studies show that the proposed entry-blurring scheme improves the application quality by up to 168% over the baseline and successfully achieves the desired accuracy. The latency overhead of our method is negligible and the increase on area and power cost can be minimized to 6% and 8%, respectively.