TENSURE: Fuzzing Sparse Tensor Compilers (Registered Report) (www.ndss-symposium.org)

Researchers from Virginia Tech have introduced TENSURE, the first specialized blackbox fuzzing framework aimed at enhancing the testing of Sparse Tensor Compilers (STCs). As STCs become essential in optimizing machine learning and high-dimensional data analytics, their susceptibility to subtle bugs due to complex control flows poses a significant risk. Traditional testing methods have fallen short, especially when faced with the unique requirements of STCs, which involve intricate tensor operations and storage formats. TENSURE addresses these challenges by utilizing Einstein Summation notation to generate complex tensor contractions that reveal corner cases and potential faults in STC code generation. Its innovative constraint-based generation algorithm achieves a remarkable semantic validity rate of 100%, a stark improvement over the mere 3.3% effectiveness observed in existing grammar-based fuzzers. Additionally, TENSURE employs semantic-preserving mutation operators to facilitate metamorphic testing without requiring a trusted baseline. Evaluations conducted on leading systems like TACO and Finch uncovered numerous vulnerabilities, emphasizing the urgent need for targeted testing solutions in the sparse compilation landscape and reinforcing the importance of rigorous validation in advancing AI and machine learning technologies.