Detecting Manipulated Input Data via Distortion Analysis

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for classifying input data in digital imaging devices, such as those used in automotive applications, are vulnerable to manipulation, leading to incorrect classification results that can be difficult to detect, especially when the manipulated data appears similar to authentic data.

Innovation Solution

A computer-implemented method that uses class-specific reference data based on distortions applied to input data to determine if it has been manipulated, involving the creation of distorted data sets, classification of these sets, and comparison using similarity measures like cosine similarity to identify manipulated input data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manipulated input data is used for classification, then wrong classification results are obtained, but the manipulated data appears similar to authentic data making detection difficult

Engineering Contradiction:
Improveclassification accuracyVSAvoiddetection difficulty
Core Design Contradiction:
Measurement precisionVSDifficulty of detecting and measuring

Solution Approach 1:

The patent applies distortions to training data in advance to generate distorted training data sets, which are then used to train the classification system. This preliminary action enables the system to learn robust features that are invariant to common distortions, thereby improving detection capability before actual classification occurs.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces distorted versions of input data as intermediaries to detect manipulated data. By comparing the classification results of original and distorted data, the system can identify inconsistencies that indicate manipulation, using the distorted data as a mediator to reveal hidden anomalies.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If multiple distortions are applied to detect manipulated data, then detection robustness is improved, but computational complexity increases

Engineering Contradiction:
Improvedetection robustnessVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the detection process into multiple independent distortion operations applied sequentially to the input data. Each distortion is applied separately and its classification result is recorded, allowing the system to process distortions in manageable segments rather than as a single complex operation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent changes parameters of the input data by applying various distortion transformations (such as geometric transformations, color space transformations, etc.). By systematically varying these parameters through multiple distortions, the system achieves robust detection while maintaining manageable computational complexity through parameter-based transformations.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3757867A1Method and system for determining whether input data to be classified is manipulated input data
Publication Date: 2020.12.30 APTIV TECHNOLOGIES AG
  • EP3757867A1 patent drawingFigure 1~2
  • EP3757867A1 patent drawingFigure 3
  • EP3757867A1 patent drawingFigure 4

AI summary

A computer implemented method for determining whether input data to be classified into one of a plurality of classes is manipulated input data comprises the following steps carried out by computer hardware components: providing class-specific reference data for each class based on at least one distortion; applying the at least one distortion to the input data to obtain at least one distorted input data set; classifying the input data to obtain a reference class; classifying the at least one distorted input data set to obtain at least one distorted classification result; and determining whether the input data is manipulated input data based on the class-specific reference data for the reference class and based on the at least one distorted classification result.