Decentralized Machine Learning Trust Scores for Attack Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional decentralized machine learning systems are vulnerable to adversarial attacks, particularly in federated learning systems, due to the assumption of a central parameter aggregator's genuineness, leading to a single point of failure and lack of robustness in detecting and isolating malicious nodes.
Innovation Solution
A decentralized ML system implemented on a blockchain network where each data processing node shares learning parameters, determines similarity, and publishes cryptographic hash values and trust scores on a distributed ledger, dynamically updating node weightages to exclude malicious nodes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a central parameter aggregator is used in federated learning, then the system can merge learning parameters efficiently, but the system becomes vulnerable to adversarial attacks and creates a single point of failure
Solution Approach 1:
The patent extracts the central parameter aggregator from the federated learning system and replaces it with a decentralized consensus mechanism. Each node independently validates and merges parameters through cryptographic proofs and consensus protocols, eliminating the single point of failure while maintaining merging functionality through distributed computation.
Solution Approach 2:
The patent segments the centralized parameter aggregation function into multiple distributed nodes. Each node performs partial aggregation and validation, with the overall merging process distributed across the network through consensus mechanisms, thereby improving reliability while preserving productivity.
2Device complexity
If decentralized learning is implemented across multiple nodes, then processing burden is reduced, but the system becomes vulnerable to malicious nodes injecting poisoned updates
Solution Approach 1:
The patent implements feedback mechanisms where each node validates incoming parameter updates from other nodes before incorporation. The system monitors update patterns, detects anomalies through statistical analysis and cryptographic verification, and provides feedback to exclude malicious updates, thereby maintaining decentralized processing while defending against poisoned updates.
Solution Approach 2:
The patent applies preliminary anti-action by implementing pre-validation checks on parameter updates before they are merged into the global model. Cryptographic signatures, consistency checks, and anomaly detection are performed in advance to prevent malicious updates from affecting the learning process, countering threats before they can cause harm.
3Reliability
If trust scores and node weightages are dynamically updated, then malicious nodes can be isolated, but computational overhead increases
Solution Approach 1:
The patent applies partial action by implementing trust score updates only for nodes exhibiting suspicious behavior patterns rather than continuously updating all nodes. Computational resources are allocated selectively based on detected threats, reducing overall overhead while maintaining the ability to isolate malicious nodes when needed.
Solution Approach 2:
The patent implements preliminary action by pre-computing cryptographic signatures and validation metrics for parameter updates. Trust scores are updated incrementally based on accumulated evidence rather than requiring full recomputation, reducing computational overhead while maintaining reliable malicious node detection and isolation.
Data Source
AI summary
A system and a method for detecting and defending against adversarial attacks in decentralized learning models are described. The method comprises obtaining a learning parameter for determining a reference cryptographic hash value and a similarity between the data processing nodes (102). A cryptographic hash value is determined for each data processing node (102) based on the learning parameter. The trust score of each data processing node (102) is updated based on matching of the cryptographic hash value with the reference cryptographic hash value. The learning parameter of each data processing node (102) is merged to obtain a merged learning parameter based on the trust score. The merged learning parameter is provided to the data processing nodes (102) to be used for training the machine learning models.