From e609cc2a2c315699ae6a2920457743fba8399558 Mon Sep 17 00:00:00 2001 From: shiva108 Date: Sun, 4 Jan 2026 00:21:43 +0100 Subject: [PATCH] docs: Refine formatting of numbered lists and add a blank line for readability. --- docs/Chapter_20_Model_Theft_and_Membership_Inference.md | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/docs/Chapter_20_Model_Theft_and_Membership_Inference.md b/docs/Chapter_20_Model_Theft_and_Membership_Inference.md index 30ecbf9..271d47c 100644 --- a/docs/Chapter_20_Model_Theft_and_Membership_Inference.md +++ b/docs/Chapter_20_Model_Theft_and_Membership_Inference.md @@ -942,9 +942,9 @@ PRIVACY VIOLATION: Model leaks training data membership ### Current Research Gaps -1. **Copyright inWeights**: Determining if a model "contains" a copyrighted work in a legal sense (substantial similarity). -2. **Machine Unlearning**: How to remove a distinct concept/person from a model cost-effectively. -3. **Watermark Robustness**: Creating watermarks that survive distillation/theft (most currently fail). +1. **Copyright inWeights**: Determining if a model "contains" a copyrighted work in a legal sense (substantial similarity). +2. **Machine Unlearning**: How to remove a distinct concept/person from a model cost-effectively. +3. **Watermark Robustness**: Creating watermarks that survive distillation/theft (most currently fail). ### Recommended Reading @@ -973,6 +973,7 @@ As models move to the edge and APIs become ubiquitous, these "grey box" attacks ## Quick Reference ### Attack Vector Summary + Attackers query the model to either learn its internal parameters (Model Theft) or determine if specific data points were used during training (Membership Inference). This exploits the model's high information retention and correlation with its training set. ### Key Detection Indicators