Thread by @ram_ssk

THREAD: When ML system fail because of adversarial manipulation, how should society expect law to respond?

I've been working with lawyers @d_obrien, @KendraSerra, @salome_viljoen to figure this out and here's how ML engineers can help

Paper: arxiv.org/abs/1810.10731

1. Benchmark your defenses and attacks - Public policy experts dont know how to prioritize the published defenses and attacks.

Invest in community benchmark like robust ML (a la @aleks_madry ) and report your results after tools like cleverhans.

2. Architect your ML systems for forensics - It will help the security people during investigation/response and help the legal folks later

3. Think about the "Dual use of adversarial ML" and civil rights - Governments can backdoor ML systems; dissidents can escape facial recognition through specially crafted eyewear.

4. The big question is "Why should adversarial ML be different from software based attacks in front of the law". The sooner we can tabulate the differences, the better position we will be in.

5. Interested in this area? Some follow ups:
- @rcalo work "Is tricking a Robot hacking?". He is my tech lawyer bae, and writes clearly!
- @jackclarkSF in his awesome ImportAI letter frequently writes about security, and has a policy section . Just piece the two yourself!

6. Finally, on a personal note, interdisciplinary research is the ultimate growth mindset exercise. You have to ask dumb questions all the time. For instance, @d_obrien spent a good half hour yesterday explaining the difference between statute and regulation. Everyone learns!