On-Device Scam Message Detection With In-Context User Guidance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for identifying scam electronic messages are ineffective on-device due to server-side encryption and lack of contextual guidance, leading to high-risk messages going undetected and unaddressed, especially in messaging applications that encrypt data on the server and in transit.
Innovation Solution
A method using machine learning (ML) and natural language processing (NLP) to analyze messages on mobile devices and desktops, providing in-context alerts and guidance for users to identify high-risk scam messages, with optional guardian support for vulnerable populations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If server-side encryption is used to protect message privacy, then user privacy is improved, but the ability to detect scam messages is worsened
Solution Approach 1:
The patent introduces an intermediary component - an on-device analysis engine that acts as a mediator between encrypted messages and user detection needs. This engine analyzes message content locally on the user's device without requiring server-side decryption, thus maintaining privacy while enabling scam detection through intermediate local processing
Solution Approach 2:
The patent segments the detection system into server-side components (which handle encrypted data) and client-side components (which perform actual analysis). By dividing the detection function across server and device, the system maintains encryption while enabling detection through distributed architecture
2Ease of operation
If binary assessment (scam/not scam) is provided by machine learning, then detection simplicity is improved, but accuracy and user safety are worsened due to probabilistic nature
Solution Approach 1:
The patent changes the output parameter from a simple binary classification to a multi-level risk score system. Instead of just scam/not scam, the ML model outputs probabilistic risk scores that indicate the likelihood of scam, allowing users to make informed decisions based on the degree of risk rather than a definitive binary label
Solution Approach 2:
The patent implements feedback mechanisms where users can report false positives or confirm scams, and this feedback is used to continuously refine and improve the ML model's accuracy over time, addressing the probabilistic nature by learning from real-world outcomes
3Difficulty of detecting and measuring
If on-device analysis is implemented to analyze encrypted messages, then scam detection capability is improved, but device complexity and processing requirements are worsened
Solution Approach 1:
The patent applies preliminary action by pre-training ML models on server infrastructure and preparing analysis engines before deployment to devices. The complex model training and data processing are performed in advance on powerful servers, leaving only inference operations for the mobile device, thus reducing on-device complexity
Solution Approach 2:
The patent substitutes complex mechanical analysis systems with optimized machine learning models that are tailored for mobile device hardware. By using specialized ML architectures and leveraging device-specific hardware accelerators, the system achieves effective scam detection without proportionally increasing device complexity
Data Source
AI summary
One embodiment provides a computer-implemented method that includes receiving, by a processor, an indication of messaging applications to monitor, specifying a risk threshold. The processor further provides a notification to at least one device. The notification includes an identified scam message from the messaging applications. The processor additionally provides step-by-step guidance on actions to take to the at least one device.