Iron Mountain InSight DXP: AI model training, security and compliance

Data privacy and security

InSight DXP prioritizes data privacy and security through robust measures:

• Data privacy and security: The processing of data is subject to the relevant data privacy and security standards following applicable regulations and industry best practices to maintain confidentiality, integrity, and controlled access.
• Encryption: The platform employs industry-standard TLS 1.2 or higher for data in transit and AES 256 encryption for data at rest. This helps protect customer data from unauthorized access.
• Exclusive deployment: AI models are deployed within Iron Mountain’s cloud infrastructure to provide exclusive access and help prevent data exposure to third-party AI or cloud vendors.
• Model behavior & controls: Our sovereign gateway and K8s isolation restrict agents from raw data, moving us from "honor-system" governance to "infrastructure-enforced" security.
• Localized training: InSight DXP utilizes only the customer’s own data for model training and validation, so that one customer’s data is never used to train models for another customer.
• Access control: The platform implements multi-factor authentication and allows users to define data access boundaries, maintaining strict control over who can access sensitive information.
• Data loss prevention: The platform leverages data loss prevention (DLP) mechanisms to identify and mask personally identifiable information (PII), while preserving both original and de-identified data copies, maintaining compliance with data privacy regulations. 
• Bias & input management: Our mandatory PII scrubbing microservice acts as a hard gate, physically masking identifiers before they reach an LLM to prevent biased processing. 
• Partner selection: Iron Mountain partners with vendors that adhere to strict data security and privacy standards and controls (technical, administrative, and contractual).
• Logging and monitoring: The platform logs and monitors AI prompts and responses to detect and help prevent exposure of sensitive data to the LLM, supporting data protection, auditability, and compliance requirements.

Ethical AI and bias mitigation

Iron Mountain employs a multi-pronged approach to promote ethical AI development and mitigate bias:

• Diverse data sources: The platform utilizes a range of data sources, including customer data, open-source data, and simulated data, so that model training is comprehensive and representative.
• Human evaluation: Multi-labeler consistency methods and SME validation are employed to maintain data quality and reduce bias in training datasets.
• Continuous monitoring: Automated systems and manual processes, including a dedicated security information and event management (SIEM) system and cyber incident response team (CIRT) team, continuously monitor anomalies and potential biases.
• Model prediction tracing: Connecting model predictions to document resources to avoid unintended results.
• Knowledge base and domain-specific rule engine: Calibrating prediction confidence scores and guiding interactions within the context of relevant documents. 
• Response filters: Constraining interactions within the context of documents, helping to  prevent unintended outputs.

Reliability and accuracy

Before production deployment, Iron Mountain rigorously tests model reliability and accuracy through: 

• Data collection and cleaning: Gathering new and diverse datasets, including recent customer data, to maintain continued relevance and reduce noise or errors.
• Rigorous testing: New models undergo both quantitative and qualitative evaluation, including accuracy metrics, response latency testing, and human-in-the-loop (HITL) review.
  • Quantitative evaluation: Calculating accuracy metrics, response latency, and batch-data stress testing.
  • Operational quality (non-HITL): We are implementing automated schema validation to reject malformed or low-confidence outputs at the code level, maintaining accuracy without over-relying on manual review.
  • Evaluation: Human labelers use human-in-the-loop (HITL) to check prediction results and submit review reports. 
  • A/B testing: Comparing the performance of new models with existing models.
• Model retraining: Utilizing fine-tuning techniques to optimize model performance.
• Thorough evaluation: Rigorous assessment and comparison drive the deployment of the optimal solution.
• Feedback loop: User feedback is actively captured and incorporated to identify areas for improvement and enhance model performance. Users can submit tickets to report bugs or performance issues.
• Industry standard accuracy metrics: Utilizing established metrics to track model performance.

Error detection and correction

Before production deployment, Iron Mountain uses automated systems and manual processes to detect and correct errors or biases in AI models, such as those previously listed. These systems include:

• Benchmark datasets: Regular evaluation using benchmark datasets helps identify errors and biases.
• Human-in-the-loop (HITL) correction: Exception based HITL flagging and correction mechanisms are used to address anomalies.