deepdrift 1.1.2

Zero-Training Hidden-State Monitoring for Robustness in Vision, Language, and Generative Models

DeepDrift

DeepDrift is a runtime neural network monitoring library based on analyzing the rate of change of hidden states (Semantic Velocity). It requires no additional training, calibrates only on normal data, and introduces <1.5% computational overhead.

---

Installation

# Install from source
git clone https://github.com/Eutonics/DeepDrift.git
cd DeepDrift
pip install .

# Or via pip (after publication)
# pip install deepdrift

---

Quick Start (ViT OOD)

from deepdrift import DeepDriftMonitor
from torchvision.models import vit_b_16
import torch

model = vit_b_16(pretrained=True)
monitor = DeepDriftMonitor(
    model,
    layer_names=['encoder.layers.5', 'encoder.layers.11'],
    pooling='cls'
)

x = torch.randn(1, 3, 224, 224)
_ = model(x)

velocities = monitor.get_spatial_velocity()
print(f"Peak velocity: {max(velocities):.4f}")

---

Use Cases

1. OOD Detection in Vision Transformers

Use DeepDriftVision for automatic calibration and Out-of-Distribution detection.

Dataset Metric AUC (DeepDrift)

---

CIFAR-100 (In) - - SVHN (OOD) Peak Velocity 0.982 LSUN (OOD) Peak Velocity 0.975

from deepdrift import DeepDriftVision

monitor = DeepDriftVision(model)
monitor.fit(train_dataloader)

diagnosis = monitor.predict(test_batch)
if diagnosis.is_anomaly:
    print("OOD detected!")

---

2. Predicting RL Agent Collapse

from deepdrift.rl import DeepDriftRL

monitor = DeepDriftRL(agent.policy, threshold=0.15)

for obs in episode:
    diag = monitor.step(obs)
    if diag.is_anomaly:
        print(f"Warning: High instability detected! Velocity: {diag.peak_velocity}")

---

3. Hallucination Detection in LLMs

from deepdrift import DeepDriftGuard

guard = DeepDriftGuard(llm_model)

for token in generation_loop:
    diag = guard(token)
    if diag.is_anomaly:
        print("Possible hallucination detected.")

---

4. Early Memorization Detection in Diffusion Models

See experiments/diffusion_memorization.py.

---

API Reference

DeepDriftMonitor

• __init__(model, layer_names, pooling, n_channels, ...)
• get_spatial_velocity()
• get_temporal_velocity()
• calibrate(dataloader)

DeepDriftVision / DeepDriftGuard

Specialized wrappers with simplified API.

---

Reproducing Experiments

git clone https://github.com/Eutonics/DeepDrift
cd DeepDrift
bash scripts/reproduce_all.sh

---

Citation

@article{evtushenko2026deepdrift,
  title={DeepDrift: Zero-Training Hidden-State Monitoring for Robustness in Vision, Language, and Generative Models},
  author={Alexey Evtushenko},
  year={2026},
  journal={arXiv preprint},
  url={https://github.com/Eutonics/DeepDrift}
}

---

License

MIT

---