{{ message }}
Repository navigation
[Feature] Add a command to save all transformed query attacks #348
Description
A command to save all transformed query attacks with their details such as the model output confidence is needed for people who work on attack detection systems especially for the black-box attacks in which they need a sequence of all queries that an attacker makes to fool the model. I'm working on algorithms to detect the attacks and I need sequences of attack queries, so a command for it could be useful.
Activity
jxmorris12 commentedon Nov 19, 2020
Hey @threemmm -- thanks for the suggestion! To be specific: you're suggesting a feature in which we save intermediate transformed texts and their scores to a file, not just the last transformation, right?
QData#348 Add --save-sequences-csv command to save all intermediate t…
threemmm commentedon Nov 19, 2020
Hey @jxmorris12 , thank you for your awesome package.
Yes, exactly. I did it by adding "--save-sequences-csv" command, and I tried to follow your structure to save them (adding a new class to logger, and save into a file by flush() at the end of the process). However, it doesn't work well with all search methods, as I return the result of "get_goal_results()" in addition to the final perturbed result in search methods to quickly meet my need.
I tested with textbugger recipe and albert-base-v2-sst2 , and the output was as the photo below
--recipe textbugger --model albert-base-v2-sst2 --save-sequences-csv --num-examples 3I'd like to hear your thoughts.
github-actions commentedon Jan 19, 2021
Stale issue message
qiyanjun commentedon Jan 22, 2021
Good enhancement to add.. Will update.
qiyanjun commentedon Aug 13, 2026
This is still a reasonable ask (and @threemmm's prototype comment shows it's mostly a new logger class), but a proper implementation needs to work across all search methods consistently, which @threemmm noted their prototype didn't fully solve. That's real design/engineering work rather than a bounded fix, so leaving this open rather than guessing at an implementation. Anyone interested in picking this up: the natural shape is a new logger (similar to the existing CSV/JSON loggers) that hooks into each search method's intermediate `get_goal_results()` calls rather than just the final result.