Duplicating Black Box Machine Learning Models
summary
Describes stealing cloud-hosted black-box ML models accessed via prediction APIs by repeatedly querying them to train a clone. Proposes GDALR (Gradient Driven Adaptive Learning Rate), a modification of gradient-descent that adapts the learning rate from gradient values to achieve lower loss and faster convergence with fewer epochs, reducing time and cost of duplication and challenging existing MLaaS countermeasures.
AI-written summary · about
chapters (11)
- 0:00Intro
- 3:10MULTI LAYER PERCEPTRON (MLP)
- 10:55Model stealing/duplication techniques
- 13:14Offline attacks
- 20:02Azure ML business model What is Azure Machine Learning
- 21:52Online attacks
- 24:08Inefficiencies with present attack methodology
- 29:00Scope for Attack optimization
- 30:10Mathematical modification to curreny attack methodology
- 32:01Experimental setup
- 32:38LOGISTIC REGRESSION
related talks
28mStealing Machine Learning Models via Prediction APIs
20mAttacking and Defending Machine Learning Applications of Public Cloud
46mIncreasing Participant Agency by Understanding the Impact of Training Data
02h 13mUnderstanding Security Threats Against Machine/Deep Learning Applications
01h 04mPredicting and optimizing the behavior of large ML models
30mFrom Black Box to a Known Quantity: How to Build Predictable, Reliable ML-based
38mBuilding robust machine learning systems
17mModel-Switching: Dealing with Fluctuating Workloads in Machine-Learning-as-a-Service