A job seeker who has been consistently rejected by automated resume screening systems reads a viral LinkedIn post describing a trick: paste a large block of relevant keywords in white-colored 12pt font at the bottom of the resume PDF. The text is invisible to human reviewers but readable by ML models extracting text from PDFs, boosting the resume's score above the rejection threshold.
A company's HR platform uses an ML resume screening tool. A job seeker reads a blog post claiming that adding certain invisible keywords (white text on white background) to a resume PDF causes ML screening tools to score the resume higher. This is an example of:
- A.B. An evasion attack - manipulating the input at inference time to cross the model's decision boundary
- B.A. A backdoor attack - the invisible text is a trigger
- C.C. A membership inference attack - the keywords signal the applicant was in the training set
- D.D. Data poisoning - adding the keywords permanently corrupts the model
Why A is correct
This is an evasion attack: the job seeker modifies their input (resume) at inference time to push the model's prediction across the decision boundary (from 'reject' to 'accept'), without any access to the model. It is not a backdoor (which requires training access). It is not membership inference (which probes training set membership). It is not data poisoning (which modifies the training pipeline).
Know someone studying for AI Security Fundamentals? Send them this one.