Multi-Model Frame Classification for Patch Attack Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models are vulnerable to patch attacks, which degrade accuracy by manipulating the certainty of object classification, making it difficult to discern between legitimate algorithmic issues and actual attacks.

Innovation Solution

Employ two distinct machine learning models (e.g., YOLO and Faster RCNN) to analyze frames, identify frames with inconsistent classification results, and determine a patch attack by comparing the certainty differences between frames using a predefined threshold.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a single machine learning model is used for classification, then the system complexity is low, but the reliability of detecting patch attacks is insufficient

Engineering Contradiction:
Improvepatch attack detection accuracyVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent combines multiple machine learning models (first ML model and second ML model) into a unified detection system. The models work together to process the same image data and compare their classification results, enabling more reliable patch attack detection through their collective decision-making capability rather than relying on a single model's judgment.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces an image processing device as an intermediary that coordinates between multiple ML models. This intermediary device collects classification results from different models, compares their certainty values, and determines whether a patch attack is present based on the discrepancies, thereby managing the complexity of multi-model integration.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If multiple machine learning models are used to compare classification results, then the reliability of patch attack detection is improved, but the device complexity increases

Engineering Contradiction:
Improvepatch attack detection accuracyVSAvoidnumber of machine learning models
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent uses multiple ML models to perform classification, which is more than the single model traditionally used. This excessive action of deploying multiple models creates redundant classification results that can be compared to detect inconsistencies caused by patch attacks, accepting the added complexity as necessary for improved security.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent changes the parameter of model diversity by selecting ML models with different architectures or training approaches. This parameter change ensures that the models process the same input differently, making their comparison more effective for detecting patch attacks that might affect all models uniformly or differently depending on the attack type.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If the threshold for determining patch attacks is set low, then the detection sensitivity is high, but false positives increase causing misidentification of legitimate users

Engineering Contradiction:
Improvepatch attack detection sensitivityVSAvoidfalse positive rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements a feedback mechanism where the detection system continuously monitors the classification results from multiple ML models and adjusts its determination based on the consistency of their outputs. The system uses the certainty values and class comparisons as feedback to dynamically assess whether a patch attack is present, rather than relying on a fixed low threshold that would cause false positives.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4618042A1Computer-readable recording medium, information processing method, and information processing device
Publication Date: 2025.09.17 FUJITSU LTD
  • EP4618042A1 patent drawingFigure 1
  • EP4618042A1 patent drawingFigure 2
  • EP4618042A1 patent drawingFigure 3

AI summary

An information processing device (100) uses a first machine learning model (110) and obtains a first classification result (111). The information processing device (100) uses a second machine learning model (120) and obtains a second classification result (121). The information processing device (100) identifies a first frame, based on the first classification result (111) and the second classification result (121). The information processing device (100) identifies a first class with a highest degree of certainty in the first classification result (111) for any frame regarded as a reference. The information processing device (100) identifies a second frame that is immediately before or after the identified first frame. The information processing device (100) determines a series of frames to be invalid when a predetermined condition is satisfied by a relationship between a first degree of certainty of the first class in the first classification result (111) for the first frame and a second degree of certainty of the first class in the first classification result (111) for the second frame.