AI System Vulnerability Assessment for Sponge Attack Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing AI systems are vulnerable to sponge attacks that increase energy consumption and runtime latency, posing a threat to critical applications like autonomous driving and industrial robots, leading to performance degradation and potential Denial-of-Service.

Innovation Solution

A framework and method to assess AI system vulnerability by adding perturbations to inputs, monitoring inference time, and using a defense model to identify and block adversarial inputs, ensuring processing times stay within predefined limits.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the AI system processes real-time data using trained models, then the system generates accurate outputs for critical applications, but the system becomes vulnerable to sponge attacks that increase energy consumption and runtime latency

Engineering Contradiction:
Improvesystem performanceVSAvoidenergy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The framework performs preliminary vulnerability assessment by adding perturbations to training data before the AI system is deployed or before processing critical real-time data. This proactive approach identifies sponge attack vulnerabilities in advance, allowing the system to be hardened against energy consumption attacks before they occur, thus preventing excessive energy usage while maintaining reliability

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The framework introduces an intermediary assessment layer between the AI model and the input data. This intermediary component evaluates the vulnerability of the model to sponge attacks by analyzing perturbed training data, and can prevent malicious inputs from reaching the model, thereby protecting the system from energy consumption increases while preserving accurate output generation

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If the AI system processes real-time data using trained models, then the system generates accurate outputs for critical applications, but the runtime latency increases under sponge attacks leading to Denial-of-Service

Engineering Contradiction:
Improvesystem performanceVSAvoidruntime latency
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The framework performs preliminary vulnerability assessment by adding perturbations to training data before the AI system is deployed or before processing critical real-time data. This proactive approach identifies sponge attack vulnerabilities in advance, allowing the system to be hardened against runtime latency attacks before they occur, thus preventing excessive processing delays while maintaining reliability

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The framework introduces an intermediary assessment layer between the AI model and the input data. This intermediary component evaluates the vulnerability of the model to sponge attacks by analyzing perturbed training data, and can prevent malicious inputs from reaching the model, thereby protecting the system from runtime latency increases while preserving accurate output generation

Inventive Principle:
Principle #24Intermediary (Mediator)

3Adaptability or versatility

If adversarial inputs are allowed to pass through the AI model, then the model can process diverse inputs, but the energy consumption and processing time increase causing performance degradation

Engineering Contradiction:
Improveinput processing capabilityVSAvoidsystem efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The framework introduces an intermediary vulnerability assessment component that sits between the input data and the AI model. This intermediary evaluates whether inputs are adversarial by analyzing perturbed training data patterns, and can filter or alert on malicious inputs before they reach the model, thus maintaining system efficiency while preserving the model's ability to process legitimate diverse inputs

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The framework extracts and removes the harmful adversarial component from inputs by identifying sponge attack patterns through vulnerability assessment. By taking out the malicious perturbations before they affect the model processing, the system maintains productivity and efficiency while still being able to handle legitimate varied inputs

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP4597373A1A framework to assess the vulnerability of an ai system and a method thereof
Publication Date: 2025.08.06 ROBERT BOSCH GMBH
  • EP4597373A1 patent drawingFigure 1
  • EP4597373A1 patent drawingFigure 2
  • EP4597373A1 patent drawingFigure 3

AI summary

The present invention proposes a framework (100) for assessing vulnerability of an AI system (10). The framework (100) comprises the AI system (10) that is in communication with a processor (11). The AI system (10) is configured to process an input via the AI Model (M) and give an output within pre-defined functional characteristics. The processor (11) is configured to generate a perturbed input x'' with minimum sparsity, which when fed as input to the AI Model (M) gives an output within pre-defined functional characteristics . The processor (11) records the inference time (T) of the AI Model (M) for processing x". The vulnerability of the AI Model (M) is assessed based on analysis of real-time inputs processed and the inference time (T).