Embedded Files
As illustrated in the paper, we rely on the attention scores to locate sensitive tokens in an input prompt, and then apply several semantically-equivalent transformations (derived and extended from the CCTest’s codebase) to iteratively mutate those sensitive tokens until generating AEs.
Transformations implementation
Transformations implementation
We use tree-sitter to parse the code. You can refer to tree-sitter. https://tree-sitter.github.io/tree-sitter/
We reuse the CCTest to apply the semantically-equivalent transformations on Python. Readers can first visit its codebase for tutorials. https://sites.google.com/view/cctest-info/main-page
Additional mutations
Additional mutations
Besides the default mutations CCTest offers, we add the following two new mutation passes in our experiments.
Pass 1
Pass 1
This pass first collects all the potential variables and replaces them with equal expressions
We give a simplified code example on the right side.
Pass 2
Pass 2
This pass first analyses the usage of while/for, and then changes them interactively if applicable.
Noted that not all the while statements can be replaced by for statements.
We give a simplified code example on the right side.
AE Example
AE Example
Page updated
Google Sites
Report abuse