Observability
& Security Tooling
for Latent Space Applications, Language Models, etc
<warning> Improper Monitoring of Language Models & Latent Space Applications poses an Existential Risk via Universal, Transferable, and Infinitely Generatable Attack-Strings; Secure your Environments ASAP </warning>
Vulnerabilities exist in core depenencies of Large Language Models (Transformers, Common Crawl, etc). These attack strings are infinitely generatable, transferable, and can be automatically templated; allowing for easy localization and utilization by malicious actors.
We're fixing the 'Log4j of LLMs' one step at a time; executive summary available
Latent Space Tools help conceptualize, visualize, and subsequently operationalize the necessary architecture and software components for secure LLM Deployment & Monitoring.
Open-Licensure & Distribution
Latent Space Tools are made available under the Apache 2 license via Github
Key Components
Input Pre-Processing
1) Prompt Injection Detection & Mitigation
2) Service Denial & Performance Monitoring
Data Enrichment, Monitoring & Clustering
3) Topic / Sentiment Modeling x Vector Comparisons & Cluster Definition
Output Post-Processing
4) Attack Mitigation, Appending (Un)Certainty & Response Non-Conformity
Output Forecasting
5) Heatmaps x Dimensionality Drift via Conformal Prediction Intervals
Note: Actively developing models designed as additional pre-processing to differentiate attack strings vs parameterized URLs; also looking to develop membership and attribute inference attacks as pipelines to affect point-forward GDPR compliant 'forgetting' for DNNs utilizing open-source tools like WeightWatcher.ai for layer-specific validation.
Architectural Overview
based on A16Z's Reference Architecture; now with grounding
more details available on GitHub
Core Concepts
N-Dimensional Drift:
Given that a latent space generally represents a reduced dimensionality compared to the feature space, we expect the 'aggregate' dimensions to move around more than their component parts.
That said, the chosen dimensions should represent meaningful metrics worth monitoring. Hence, the importance of conceptualizing, monitoring, and forecasting changes to those values.
Conformal Prediction:
Latent Space Tools extensively leverage the concept of conformal prediction; whereby previous outputs better predict future outputs than do Bayesian priors or assumptions.