Machine learning's increasing power and accessibility, exemplified by complex tasks like panoptic segmentation, also introduces significant security vulnerabilities.
Transfer learning allows developers to repurpose large pre-trained teacher models for specific tasks with minimal data and compute by fine-tuning a new student model.
The core benefits of transfer learning, such as knowledge transfer and minimal training, directly create attack vectors for adversaries.
Adversarial examples can fool models with subtle input changes (evasion), while poisoned data can insert hidden backdoors, with both risks amplified by transfer learning.
Before writing code, perform a thorough risk assessment by defining security requirements, evaluating resource availability, and conducting threat modeling for your specific use case.
Mitigate risks by choosing transparent and trustworthy teacher models and using robust models hardened through techniques like adversarial training.
Strengthen your student model by fine-tuning all layers to diverge from the teacher model, using backdoor detection, and performing continuous stress testing.
Practical resources like the Adversarial Robustness Toolbox for developers and security principles from the National Cybersecurity Center can help you build more secure ML systems.
24:23Balázs Kiss
27:02Mirko Ross
48:01Jan Zawadzki
23:24Keno Dreßel
23:50Aditi Godbole
29:00Alex Soto
30:36Mackenzie Jackson
27:26Cassie Kozyrkov
.gif?w=240&auto=compress,format)
Jobs that call for the skills explored in this talk.
Understanding Recruitment Group
Barcelona, Spain