Advertising Fraud Detection Using Client Publisher Feature Extraction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional techniques find it difficult to detect advertising fraud when attackers use a distributed manner with multiple clients and publisher sites, making it challenging to differentiate between legitimate and fraudulent advertising requests based on burstiness.

Innovation Solution

A detector that extracts client and publisher information from advertising requests, calculates feature amounts using appearance frequencies, and generates a determiner to differentiate between benign and malignant requests, utilizing learning logs to assign labels and determine the legitimacy of advertising requests.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional burstiness-based detection techniques are used, then detection simplicity is maintained, but detection accuracy deteriorates when attackers use distributed methods with multiple clients and publisher sites

Engineering Contradiction:
Improvedetection accuracyVSAvoiddetection method complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the detection approach by dividing advertising requests into distinct categories (first advertising requests from known malicious sources, second advertising requests from unknown sources) and applying different detection strategies to each segment. This segmentation enables targeted analysis that improves detection accuracy while managing complexity through structured classification.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements preliminary action by pre-acquiring information about malicious advertising requests and storing them in a database before actual detection occurs. This pre-processing step creates a reference library of known fraudulent patterns, enabling faster and more accurate detection during operational phases without increasing real-time computational complexity.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If distributed advertising fraud methods are used by attackers, then fraud resilience improves, but detectability deteriorates

Engineering Contradiction:
Improvefraud detection reliabilityVSAvoidfraud detection difficulty
Core Design Contradiction:
ReliabilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent merges multiple detection dimensions by combining analysis of request source information, publisher information, temporal patterns, and behavioral characteristics into a unified detection framework. This multi-dimensional approach increases detection reliability by examining fraud from multiple angles simultaneously, making it harder for distributed fraud schemes to evade detection.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent implements feedback mechanisms by continuously analyzing advertising request patterns and using detection results to refine the database of malicious request information. This feedback loop enables the system to adapt to evolving fraud techniques, improving reliability over time as the system learns from new fraud patterns while maintaining organized complexity through structured data management.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3839780B1Detector, detection method, and detection program
Publication Date: 2024.10.02 NIPPON TELEGRAPH & TELEPHONE CORP
  • EP3839780B1 patent drawingFigure 1
  • EP3839780B1 patent drawingFigure 2~3(b)
  • EP3839780B1 patent drawingFigure 4

AI summary

A feature extraction unit (11) extracts, from an advertising request to view an online advertisement, client information on a client as a transmission source of the advertising request and publisher information on a website of a publisher who displays advertising, and calculates a predetermined feature amount using the client information and the publisher information with respect to a plurality of advertising requests including at least a benign advertising request, and a determiner generation unit (12) generates a determiner (13) that determines whether an advertising request is malignant or not by using the calculated feature amount.