AI model protection

Your model runs everywhere.
Nobody can steal it.

AI models represent years of training and millions in compute. eCora keeps weights and architecture encrypted during inference — even when running inside a customer's cloud.

get started how it works

eCora — model protection

host memory visibility

any

ML framework supported

attested at inference

Weights encrypted at rest and in memory

Architecture and config sealed inside container

Memory dump reveals only ciphertext

Hardware attestation per inference session

the challenge

Deploying a model means handing over the most valuable thing you've built.

Any host with memory access can extract model weights during inference and reconstruct your architecture. Cloud administrators, infrastructure staff, and malicious tenants all have this access by default — without any log entry.

Model weights readable from host memory during inference

Container filesystem exposes architecture and configuration

No way to prove the model hasn't been modified or backdoored

with eCora

Outputs come out. The model stays inside.

eCora seals your model and inference server inside a hardware-encrypted container before it leaves your machine. Weights are encrypted in memory during inference by the CPU itself — physically impossible to extract, even with root access to the host.

Weights encrypted at rest and in memory — extraction is physically impossible

Architecture and config sealed inside the container

Hardware attestation proves the model is genuine at every session

the process

Package, seal, and deploy to any customer — without exposure.

package

Containerize your model and inference server

Package your model and serving layer as a standard Docker container. PyTorch, ONNX, TensorRT, vLLM — any framework that containerizes works. No changes to the model code or serving logic.

seal

Encrypt weights and architecture locally

Run ecora seal on your model container. Every layer — including the model weights — is encrypted with hardware-bound keys on your machine. The sealed image contains no accessible plaintext.

deploy

Serve inference anywhere, stay in control

Customers run inference against your sealed model. Prompts go in, completions come out. The model weights never leave the enclave in plaintext. Revoke access instantly from your dashboard.

threat coverage

What eCora Protects Against

Threat

Protection

Technical Details

Model Weight Extraction

Weights encrypted in memory

AES-256-GCM encryption with TEE-sealed keys

Model Architecture Theft

Architecture hidden in enclave

Code execution isolated in hardware-enforced container

Training Data Leakage

Inversion attacks blocked

Anomaly detection with automatic query throttling

Cloud Admin Access

Hardware attestation prevents access

Remote attestation with TPM-rooted chain of trust

API Scraping

Rate limiting and pattern analysis

Abuse detection with automatic blocking

Side-Channel Attacks

TEE hardware prevents attacks

CPU-level isolation with cache partitioning

capabilities

Your IP stays yours — wherever your model runs.

Weight protection

Model weights are encrypted at rest and in memory during inference. No memory dump, regardless of host privileges, can reconstruct your model.

Framework-agnostic

Works with any ML framework that runs in a container. No runtime modifications, no custom serving layer, no framework-specific integration work.

Inference attestation

Each inference session is backed by hardware proof that the model is genuine and unmodified. Customers get a trust signal; you get audit evidence.

Revocable deployment

Stop any customer's inference access instantly from the dashboard. No model transfer, no redeployment, no waiting. Access ends the moment you revoke.

the eCora platform

AI model protection is one example of how eCora applies hardware-backed identity across your environment — currently covering people, infrastructure, software, and ops — with more to come. One platform. One control plane.

see the full platform