![SecureLoop: Boosting Performance and Data Security for Neural Network Accelerators SecureLoop: Boosting Performance and Data Security for Neural Network Accelerators](https://aiheadliner.com/wp-content/uploads/https://news.mit.edu/sites/default/files/images/202310/MIT-SecureLoop-01-press.jpg)
**SecureLoop: Optimizing Design for Deep Neural Network Accelerators with Data Security**
MIT researchers have developed a search engine called SecureLoop that efficiently identifies the best designs for deep neural network accelerators while maintaining data security. These accelerators are specialized hardware components used to process the massive amounts of data required for computationally intensive machine-learning applications. However, incorporating cryptographic operations to protect data from attacks complicates the design process. By considering the impact of encryption and authentication measures on performance and energy usage, SecureLoop helps engineers obtain the optimal design for an accelerator tailored to their specific neural network and machine-learning task.
SecureLoop outperforms conventional scheduling techniques by improving accelerator performance while keeping data protected. It is especially useful for demanding AI applications, such as autonomous driving and medical image classification, that require high speed and performance while ensuring data security.
The researchers, led by Kyungmi Lee, a graduate student in electrical engineering and computer science, collaborated with MIT professors Joel Emer, Mengjia Yan, and Anantha Chandrakasan to develop SecureLoop. The research will be presented at the IEEE/ACM International Symposium on Microarchitecture.
**Secure Acceleration for Deep Neural Networks**
A deep neural network consists of interconnected nodes that process data in multiple layers. To optimize processing efficiency, deep neural network accelerators use an array of computational units to parallelize operations in each layer. However, storing data off-chip makes them vulnerable to attackers who could steal or alter the information. Adding authenticated encryption to the accelerator helps protect data. Encryption scrambles the data using a secret key, while authentication assigns a cryptographic hash to each data chunk and stores it in off-chip memory.
However, the sizes of authentication blocks and data tiles can pose challenges. Multiple tiles can be included in one authentication block, or a tile can be split between two blocks. This mismatch can lead to energy inefficiency and computational costs.
**An Efficient Search Engine: SecureLoop**
MIT researchers enhanced an existing search engine called Timeloop to create SecureLoop. They developed a model that accurately accounted for the additional computation required for encryption and authentication. By reformulating the search problem into a simple mathematical expression, SecureLoop efficiently finds the ideal authentication block size, reducing the need for exhaustive searches. The search engine also incorporates heuristic techniques to maximize the performance of the entire deep neural network.
SecureLoop generates an accelerator schedule that includes the optimal data tiling strategy and authentication block size for a specific neural network. In simulations, SecureLoop identified schedules that were up to 33.2% faster and exhibited 50.2% better energy delay product than methods that did not consider security.
The researchers also used SecureLoop to explore how the design space for accelerators changes when security is considered. They found that allocating more space for the cryptographic engine and sacrificing some on-chip memory can improve performance.
In the future, the researchers plan to use SecureLoop for designing accelerators resilient to side-channel attacks and extend its applications to other computation types. This work is funded by Samsung Electronics and the Korea Foundation for Advanced Studies.
**Conclusion**
SecureLoop offers an efficient solution for optimizing the design of deep neural network accelerators while ensuring data security. Its capabilities make it an invaluable tool for engineers working on AI applications. With SecureLoop, they can improve performance and energy efficiency while keeping sensitive user data safe.
Note: Image source – https://news.mit.edu/sites/default/files/styles/news_article__cover_image__original/public/images/202310/MIT-SecureLoop-01-press.jpg?itok=R23pWWII