Edge Cluster Resampling for AI Data Poisoning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Edge computing environments face performance constraints and security threats due to data transfer inefficiencies and adversarial data poisoning, which affect the stability and reliability of AI models in distributed systems.
Innovation Solution
Implementing a lossy compression algorithm in edge devices to generate compressed encodings of numeric values from IoT units, which are then used to create surrogates for training AI models, and employing clustering algorithms to identify and filter out adversarial data, thereby improving data efficiency and model robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is transmitted from edge devices to cloud computing resources, then AI model training can be performed, but data communication overhead increases and system performance degrades
Solution Approach 1:
The patent extracts only the essential features from raw sensor data by computing statistical summaries (mean, variance, min, max) and transmitting these compressed representations to the cloud instead of raw data. This extraction principle reduces communication overhead while preserving sufficient information for effective AI model training.
Solution Approach 2:
The patent segments the data processing workflow into edge-side preprocessing (feature extraction, compression) and cloud-side model training. By dividing the system into these functional segments, the patent enables efficient local data handling while maintaining centralized model development capabilities.
2Measurement precision
If raw data is transmitted to the cloud for AI training, then model accuracy can be maintained, but communication bandwidth consumption increases
Solution Approach 1:
The patent transforms raw sensor data into different parameter representations (statistical moments, compressed encodings) that occupy less communication bandwidth. By changing the data parameters from raw values to aggregated statistics, the system maintains training effectiveness while reducing transmission volume by approximately 100x.
Solution Approach 2:
The patent creates compressed representations (copies) of the original data that capture essential patterns and distributions. These compressed copies include statistical summaries and clustered representations that preserve the information needed for model training while occupying minimal bandwidth.
3Adaptability or versatility
If all data is used for AI training, then comprehensive model learning is achieved, but adversarial data poisoning affects model robustness
Solution Approach 1:
The patent performs preliminary data processing and filtering at the edge device before data leaves the local system. By pre-computing statistical features and applying local validation rules, the system prepares data in a way that reduces the impact of adversarial inputs before they reach the training pipeline.
Solution Approach 2:
The patent introduces statistical summaries and compressed representations as intermediary layers between raw sensor data and the AI training process. These intermediaries act as buffers that obscure the direct mapping between raw inputs and model parameters, making it more difficult for adversarial data to directly poison the training process.
Data Source
AI summary
A method for alleviating data poisoning in an edge computing resource includes receiving a numeric value from an Internet of Things (IoT) unit and associating the numeric value with a cluster selected from a plurality of clusters in accordance with a suitable clustering algorithm such as a k-means clustering algorithm. In at least some embodiments, the numeric value comprises a poisoned numeric value including an adversarial component injected by an adversary to negatively impact a trained model of a cloud-based artificial intelligence engine. Rather than permitting the injected adversarial component to corrupt the AI engine, a cluster with which the numeric value is associated is sampled in accordance with a probability distribution of the cluster to obtain a surrogate for the poisoned numeric value. The surrogate may then be provided as an input to an inference module of the AI engine to generate a prediction.