DEMO 2: Credit Fraud Security Assessment

List available targets

list targets

Interact with creditfraud

interact creditfraud

List the available attacks (having already issued load art)

list attacks

Scan the target across attacks (50 iterations each, enable logging)

scan -n 50 --log

Save the results of the scan (note the output filename that you should use below)

save

With the --log, every model query has been logged. What's in the log? Replace <filename> with the output .json file you observed from the save command.

!tail -n 1000 <filename>

Let's take a look at a fraudulent transaction, sample_index=6899 of the data that is included in the target definition

predict -i 6899

Let's scan again with 3 iterations, but targeting only sample_index=6899 using hop_skip_jump, with the --verbose flag set.

scan -n 3 --set sample_index=6899 --attack hop_skip_jump --verbose

What is the feature vector from the last attack in the scan?

show sample -r

What was the original sample? How close is it to the perturbed input?

show sample

(Exercise for the reader) Can you use set and run to find an attack that uses the least number of queries that causes sample_index=6899 to appear benign?