An Integrated Approach to Securing AI

An Integrated Approach to Securing AI

An article by Devyani Biswal, Senior Methodology Architect, Privacy Analytics

As artificial intelligence (AI) continues to expand across various sectors, securing these systems against potential threats has become increasingly important. Our previous article, A Blueprint for Defensible AI Systems, highlighted the need for a quality management system for AI. The critical aspects of AI system design and mitigation strategies can address security threats both within and outside the system.

AI systems are susceptible to various attacks that can compromise their functionality, reliability, and trustworthiness. The NIST Adversarial Machine Learning report and guidelines from the OWASP Machine Learning Security Top Ten provide valuable insights into addressing AI vulnerabilities, particularly at the intersection of data protection and AI security. Effective AI security frameworks are essential to safeguard these systems against potential threats.

Data protection as a defense mechanism

Effective data protection strategies, such as anonymization, play an important role in mitigating AI threats. Anonymization techniques ensure that data cannot be traced back to individuals, reducing the risk of re-identification in the event of a security breach. This approach is particularly effective at limiting the scope of model inversion attacks , which attempt to extract sensitive information from the model, with regards to extracting details about data subjects.

As the OWASP Machine Learning Security Top Ten outlines, however, many other AI security threats can put data derived from people at risk. Ultimately, as “end users” of data, AI systems change the threat landscape by introducing more risks than if the end user were human. Unlike human recipients, AI systems can potentially infer other sensitive information from anonymized datasets beyond identities, and at a much larger scale. Extending defensive needs to include AI security threats is therefore becoming a critical feature of defensible AI.

Enhance your data protection with differential privacy

Find out how differential privacy and risk metrics are combined to better protect data. Read this 7-page white paper.

Comprehensive strategies for AI security

To effectively protect AI systems, we need to implement a combination of strategies addressing the various classes of issues identified by NIST: availability breakdown, integrity violations, privacy breach, and abuse violations. These strategies can include continuous monitoring and ethical hacking practices that proactively test and secure AI systems against potential vulnerabilities. By considering these classes of issues and developing comprehensive solutions around them, organizations can better safeguard their AI systems and data.

The NIST Adversarial Machine Learning report categorizes risks into four main areas, each requiring targeted mitigation strategies (examples of which are shown in the image below):

Availability breakdown: disrupting the performance of an AI model at deployment. These attacks can occur through:

  • Data poisoning: controlling a fraction of the training set to corrupt the model.
  • Model poisoning: manipulating the model parameters directly.
  • Energy-latency attacks: using query access to degrade performance.

Integrity violations: targeting the accuracy of an AI model’s output, resulting in incorrect predictions. These can include:

  • Evasion attacks: modifying test samples to create adversarial examples that mislead the model while appearing normal to humans.
  • Poisoning attacks: similar to the previous, this time corrupting the model during training to induce incorrect outputs.

Privacy breach: extracting sensitive information from the training data or the model itself. Types of privacy attacks include:

  • Data privacy attacks: inferring content or features of the training data.
  • Membership inference attacks: determining if specific data was part of the training set.
  • Model privacy attacks: extracting detailed information about the model.

Abuse violations: repurposing the system’s intended use for other objectives through indirect prompt injection. This can include:

  • Fraud: exploiting the system for financial gain.
  • Malware: spreading malicious software.
  • Manipulation: influencing outcomes or behavior for malicious purposes.

Conclusions

As AI continues to evolve, the importance of structured frameworks and standards for security becomes increasingly critical. The ISO/IEC 42001 AI Management System and the NIST AI Risk Management Framework provide comprehensive guidance for ensuring responsible and defensible AI deployments. By integrating these standards into data governance practices and robust data infrastructure, organizations can enhance data protection, scalability, and decision-making accuracy.

The future of AI governance lies in continuous improvement, ethical considerations, and the seamless integration of AI and privacy solutions. Implementing an integrated approach to mitigating AI attacks will help secure AI systems and protect sensitive data in an ever-evolving threat landscape. Contact us when you want to learn more and see how our advisory or consulting services can help you drive trustworthy insights at scale.

Archiving / Destroying

Are you unleashing the full value of data you retain?

Your Challenges

Do you need help...

OUR SOLUTION

Value Retention

Client Success

Client: Comcast

Situation: California’s Consumer Privacy Act inspired Comcast to evolve the way in which they protect the privacy of customers who consent to share personal information with them.

Evaluating

Are you achieving intended outcomes from data?

Your Challenge

Do you need help...

OUR SOLUTION

Unbiased Results

Client Success

Client: Integrate.ai

Situation: Integrate.ai’s AI-powered tech helps clients improve their online experience by sharing signals about website visitor intent. They wanted to ensure privacy remained fully protected within the machine learning / AI context that produces these signals.

Accessing

Do the right people have the right data?

Your Challenges

Do you need help...

OUR SOLUTION

Usable and Reusable Data

Client Success

Client: Novartis

Situation: Novartis’ digital transformation in drug R&D drives their need to maximize value from vast stores of clinical study data for critical internal research enabled by their data42 platform.

 

Maintaining

Are you empowering people to safely leverage trusted data?

Your Challenges

Do you need help...

OUR SOLUTION

Security / compliance efficiency

CLIENT SUCCESS

Client: ASCO’s CancerLinQ

Situation: CancerLinQ™, a subsidiary of American Society of Clinical Oncology, is a rapid learning healthcare system that helps oncologists aggregate and analyze data on cancer patients to improve care. To achieve this goal, they must de-identify patient data provided by subscribing practices across the U.S.

 

Acquiring / Collecting

Are you acquiring the right data? Do you have appropriate consent?

Your Challenge

Do you need help...

OUR SOLUTIONS

Consent / Contracting strategy

Client Success

Client: IQVIA

Situation: Needed to ensure the primary market research process was fully compliant with internal policies and regulations such as GDPR. 

 

Planning

Are You Effectively Planning for Success?

Your Challenges

Do you need help...

OUR SOLUTION

Build privacy in by design

Client Success

Client: Nuance

Situation: Needed to enable AI-driven product innovation with a defensible governance program for the safe and responsible use
of voice-to-text data under Shrems II.

 

Join the next 5 Safes Data Privacy webinar

This course runs on the 2nd Wednesday of every month, at 11 a.m. ET (45 mins). Click the button to register and select the date that works best for you.