Sensor Data Classification Using Repeated Noisy Label Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks are susceptible to data poisoning attacks, particularly label-flipping attacks, which corrupt training data and degrade classifier performance, and existing defenses are inadequate for robustness against such adversarial manipulations.
Innovation Solution
A method involving a base classifier with an encoder and a classifier part, where the encoder is fixed and the classifier is re-trained multiple times with added label noise, allowing robust predictions by determining the most frequent label among multiple training iterations, without applying randomization to the input data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data poisoning attacks corrupt training labels, then classifier performance degrades, but robustness against such attacks is insufficient with existing defenses
Solution Approach 1:
The method performs preliminary actions by repeatedly training the classifier with artificially added label noise before deployment. This pre-hardening process prepares the classifier to resist adversarial label corruption in production, addressing the robustness issue proactively rather than reactively
Solution Approach 2:
The method applies preliminary anti-action by introducing label noise during training that counteracts adversarial label flipping attacks. The artificial noise creates opposing forces that prevent the classifier from being easily manipulated by adversarial examples, thereby improving reliability against harmful factors
2Measurement precision
If training data is collected to improve machine learning quality, then classification performance improves, but corruption from faulty sensors or adversarial attacks degrades the trained product
Solution Approach 1:
The method converts the harmful effect of label noise into a beneficial training mechanism. By deliberately adding noise during training, the classifier learns to ignore corrupted labels, transforming what would normally degrade performance into a robustness-enhancing feature that maintains classification accuracy despite data corruption
Solution Approach 2:
The method changes the training parameter of label noise from zero (clean training) to a controlled non-zero value (noisy training). This parameter change allows the classifier to adapt to imperfect data conditions, maintaining measurement precision even when training data integrity is compromised
3Reliability
If repeated training with label noise is performed to improve robustness, then resistance to adversarial attacks improves, but computational cost increases
Solution Approach 1:
The method applies partial action by selectively adding label noise only to certain training samples rather than uniformly to all data. This partial application of noise achieves robustness improvement with reduced computational overhead compared to comprehensive noisy training, balancing reliability gains with training efficiency
Solution Approach 2:
The method uses periodic action by alternating between clean and noisy training epochs. This periodic alternation allows the classifier to learn from both clean and corrupted data patterns, achieving robustness without continuously incurring the full computational cost of noisy training, thereby improving productivity
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Some embodiments are directed to a method (500) to classify sensor data with improved robustness against label noise. A predicted label may be computed for a novel input with improved robustness against label noise by estimating a label which is most likely under repeated application of a base training function to the training labels incorporating noise according to a noise level and subsequent application of a base classifier configured according to the base prediction function to the novel input.