Five-Layer Anti-Cheat Infrastructure Closing off cheating paths across input, code and output so scores reflect real analysis ability Input dimension Code dimension Output dimension
Banned-literal scan Scans extractor source; a hit on a sample-specific value zeroes the robustness sub-score
Renamed-input test Reruns the extractor on a renamed sample to confirm it parses structure, not the filename
Overlay / tamper test Appends data to the sample's trailer and checks output is unchanged — immune to trailer tampering
Encoding-dependence test Zeroes the encoded bytes and confirms the rule no longer matches — anchored to encoding, not plaintext
Low-weight decoy fields Easily-brute-forced fields weighted down to ~1.5%, steering toward structured analysis
reward = w₁×artifact + w₂×correctness + w₃×robustness The five layers feed into the robustness component, keeping scores tied to real analysis ability
Blocked across three dimensions Input dim. Renamed input + overlay input: confirms the extractor parses binary structure, not the filename or trailer data Code dim. Banned-literal scan: blocks hard-coded answers — a hit on a sample-specific value zeroes the robustness sub-score Output dim. Encoding-dependence + low-weight decoys: keeps detection rules anchored to encoded bytes, suppresses brute-force scoring Coverage 10 tasks implement anti-cheat mechanisms · gate thresholds 0.30–0.95 · robustness weight 5%–10%