| 摘要: | 深度神經網絡（Deep Neural Network，DNN）已成為增強安全敏感系統（如自動駕駛汽車和醫療監控）的不可或缺的工具。然而，它們在面對硬體引發的錯誤，特別是老化效應時的可靠性仍然是一個緊迫的問題。本文對老化效應，包括負偏壓溫度不穩定性（Negative Bias Temperature Instability，NBTI）、正偏壓溫度不穩定性（Postivie Bias Temperature Instability，PBTI）和熱載子注入（Hot Carrier Injection，HCI）對深度神經網絡可靠性的影響進行了全面分析，同時提出了實用的解決方案來應對這些挑戰。
我們對記憶體裝置中的老化效應對深度神經網絡可靠性的影響進行了全面的分析，包括模擬老化條件、觀察老化誘發的錯誤（Aging-induced Error，AIE）以及此誘發之錯誤（AIE）發生在DNN上所造成的影響。透過這項研究，我們發現了老化誘發的錯誤對DNN準確性造成大幅度下降。 為了解決這個問題，我們研究了兩個有效的解決方案，包括冗餘位元注入（RBI）和降低操作頻率（LOF）。RBI方法只需25%的冗餘位元即可顯著維持 DNN 效能，而LOF方法則能使用比原本多1%的時間成本，維持深度神經網絡的性能。這些解決方案為減輕老化效應所導致的準確性下降提供了實用的解決方案，從而增強了DNN在實際應用中的可靠性。
通過完整的實驗和分析，我們揭示了老化引發的錯誤，並提出了增強深度神經網絡穩固性的方法。我們除了介紹了上述提到的兩種方法來最小化老化效應對深度神經網絡準確性的影響，還提供了一些未來可能可以發展的方向，為在現實應用中保持模型完整性提供了寶貴的見解，特別是在安全關鍵和安全敏感領域。我們的研究結果強調了采取積極措施以增強深度神經網絡韌性的重要性，為未來研究提供了基礎，旨在確保DNN在面對不斷變化的挑戰時的穩定性。
;Deep Neural Networks (DNNs) have become indispensable in bolstering security-sensitive systems such as autonomous vehicles and medical monitoring. However, their reliability in the face of hardware-induced errors, particularly aging effects, remains a pressing concern. This work conducts a thorough investigation into the repercussions of aging effects, including Negative Bias Temperature Instability (NBTI), Positive Bias Temperature Instability (PBTI), and Hot Carrier Injection (HCI), on DNN reliability while proposing practical solutions to counteract these challenges.
We conduct a comprehensive analysis of aging effects within memory devices, involving simulations to replicate aging conditions, observation of aging-induced errors (AIE), and their subsequent application to DNNs. Through this examination, we identified the detrimental impact of aging-induced errors on DNN accuracy. To address this, we study effective remedies, such as redundant bit injection (RBI) and lowering operation frequency (LOF). The RBI method, requiring only 25% redundant bits, significantly sustained DNN performance, while the LOF method minimized performance degradation to just 1% timing overhead. These remedies offer practical solutions to mitigate accuracy degradation caused by aging effects, thereby enhancing the reliability of DNNs in real-world applications.
Through meticulous experimentation and analysis, we uncover the deterministic nature of AIE and propose strategies to enhance DNN resilience. In addition to introducing the two methods mentioned above to minimize the impact of aging effects on DNN accuracy, we also provide some possible future development directions, providing valuable insights for preserving model integrity in real-world applications, particularly in safety-critical and security-sensitive domains. Our findings underscore the critical importance of proactive measures to bolster DNN resilience, laying the groundwork for future research endeavors aimed at ensuring the steadfastness of DNNs in the face of evolving challenges. |
