Improvements to TRT

[monperrus]

• [ ] Build k3s architecture for running docker containers on different machines. To scale execution horizontally.
• [ ] Better compute total execution time when retaking experiments
• [ ] Adapt TRT to solve vulnerabilities at the bug level instead of the SC-level
• [ ] Add some clever way of ensuring business logic? Replaying transactions?

Detection:

• [ ] - [ ] Establish a voting system between tools (vote in/out false positives), possibly with an ensemble model

Exploit synthesis

• [ ] TAT => Transformative Attack Tool. Like TRT but for red-teaming SC, see https://github.com/ASSERT-KTH/related-work/issues/13#issuecomment-1551513385

Better benchmark:

• [ ] Enrich the benchmark to have tests for SC in the benchmark. This would help prompt exploration => fail if SC bugs are detected

[mojtaba-eshghie]

Have we thought about the guided repair {if none of the candidates abide by the business process definition of the contract)? Guiding it by the results we get from the testing (which still needs to be implemented).

[Erikcruzk]

More Improvements Ideas

• [ ] Adapt to Smartbugs 2.0
• [ ] Create compile check on a specific version
• [ ] Add config file for parsing of results
• [ ] Do not reset execution time when retaking experiments
• [ ] Automatically create tests for SC => fail if SC bugs are detected
• [ ] TAT => Transformative Attack Tool. Shows how to attach an original SC
• [ ] TDT => Transformative Detect Tool. Detection with LLMs
• [ ] Adapt TRT to fix and assess at bug level
• [ ] Build k3s architecture for running docker containers on different machines. To scale execution horizontally.
• [ ] Increase dataset. keep hidden from LLMs
• [ ] Create test harness for all contracts
• [ ] Add Token counter to detect long prompts
• [ ] Intelligent repair of long SC that exceeds prompt max length (function level repair?)
• [ ] Create unit tests

[Erikcruzk]

@GGmorello FYI. These are some of the improvements ideas we had.