Thresholds are still research defaults or implementation-specific heuristics. They should be calibrated with known clean and known backdoored benchmark models before being treated as operational decisions.
Dataset and preprocessing mismatch can cause misleading results. This is especially important for STRIP, which relies on representative inputs and entropy under input mixing.
FreeEagle is white-box, data-free, and currently supports known ResNet-family models in Mithridatium. It does not currently support arbitrary architectures or the Hugging Face wrapper.
Hugging Face support is limited to models compatible with AutoModelForImageClassification. Architecture compatibility and preprocessing can vary by model repo.
AEVA is computationally expensive. Runtime grows with the number of source-target class pairs and HSJA query settings. Smoke tests should use reduced class ranges and small query budgets.
MMBD probes only a subset of classes by default. If a backdoor target is outside the probed set, the defense may miss it.
render_summary() has detailed summary branches for MMBD, STRIP, and FreeEagle. AEVA currently relies more on the generic fallback summary.
mithridatium/service.py contains a service-style detection path that is not fully aligned with the newer CLI options, especially FreeEagle and Hugging Face details. Treat mithridatium/cli.py as the current source of truth for command-line behavior.