Georgii Mikriukov
AI Researcher · PhD (Dr.-Ing.)
Berlin, Germany
About Me
I work on verification and safety assurance of deep neural networks through internal representations, from explainable vision models to adversarial robustness of multimodal and generative systems. My primary application domain is autonomous driving: perception pipelines, object detection and segmentation, end-to-end architectures, and safety validation of production models.
Core methods: explainable AI (XAI), adversarial robustness and attack detection, and uncertainty quantification, aligned with EU AI Act, NIST AI RMF, and TC260 AI-governance requirements. Current work extends these to vision-language models and LLM-based agents, including internal-representation probing for early detection of adversarial inputs and hallucinations.
Open Source Projects
Local Concept Embeddings
Uncertainty Gating for Cost-Aware XAI
Deep Unsupervised Contrastive Hashing
Cross-Modal Hashing with Noise Robustness
Concept-Based Adversarial Attack Analysis
Local Concept Embeddings
LoCEs (Local Concept Embeddings) provide a way to analyze how DNNs represent object concepts in complex, real-world scenes. Unlike traditional global approaches, LoCEs generate sample-specific embeddings that capture both the target object and its surrounding context within a single, compact representation.
This context-aware analysis helps uncover how models encode, separate, and confuse visual concepts across diverse scenarios.
Use cases include:
- Concept Understanding – Examine how models distinguish objects and their contexts.
- Sub-Concept Discovery – Identify unlabeled variations within categories (e.g., near vs. distant car).
- Concept Confusion Detection – Detect overlaps in representations of similar categories (e.g., bus vs. truck).
- Outlier Detection – Find unusual or challenging examples in the data.
- Information Retrieval – Search for samples using LoCE-based similarity.
- Model Comparison – Compare internal feature spaces across architectures or training methods.
Uncertainty Gating for Cost-Aware XAI
A framework for studying how epistemic uncertainty relates to XAI reliability in tabular and image settings. Epistemic uncertainty, obtained from a model's native estimator or a lightweight surrogate, is used to route samples: routing between low- and high-cost explanation methods based on expected reliability, or deferring high-uncertainty samples to save computation.
Across four tabular datasets, five architectures, and four XAI methods, epistemic uncertainty shows a strong negative correlation with explanation stability and faithfulness, a finding that also generalizes to image classification.
Deep Unsupervised Contrastive Hashing
DUCH is an unsupervised cross-modal retrieval method for efficient search and retrieval of semantically related images and text in large-scale datasets, using contrastive objectives and adversarial alignment without labeled data.
Cross-Modal Hashing with Noise Robustness
CHNR extends DUCH with a noise detection module that mitigates the impact of incorrectly paired image-text correspondences in training data.
Concept-Based Adversarial Attack Analysis
Investigates how adversarial attacks manipulate DNNs at the concept level. Adversarial perturbations are shown to be linearly decomposable into a small set of shared latent vectors, with attack components exploiting target-specific directions.
Publications
Selected Publications
Local Concept Embeddings for Analysis of Concept Distributions in Vision DNN Feature Spaces
International Journal of Computer Vision (IJCV), 2025. DOI: 10.1007/s11263-025-02446-y
Georgii Mikriukov, Gesina Schwalbe, Korinna Bade.
Evaluating the Stability of Semantic Concept Representations in CNNs for Robust Explainability
World Conference on Explainable Artificial Intelligence (xAI 2023). DOI: 10.1007/978-3-031-44067-0_26
Georgii Mikriukov, Gesina Schwalbe, Christian Hellert, Korinna Bade.
Best Industry Paper Award, xAI 2023
Unsupervised Contrastive Hashing for Cross-Modal Retrieval in Remote Sensing
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2022). DOI: 10.1109/ICASSP43922.2022.9746251
Georgii Mikriukov, Mahdyar Ravanbakhsh, Begüm Demir.
Explaining, Verifying, and Aligning Semantic Hierarchies in Vision-Language Model Embeddings
International Joint Conference on Artificial Intelligence (IJCAI 2026). DOI: 10.48550/arXiv.2603.26798
Gesina Schwalbe, Mert Keser, Moritz Bayerkuhnlein, Edgar Heinert, Annika Mütze, Marvin Keller, Sparsh Tiwari, Georgii Mikriukov, Diedrich Wolter, Jae Hee Lee, Matthias Rottmann.
Additional Publications
An Unsupervised Cross-Modal Hashing Method Robust to Noisy Training Image-Text Correspondences in Remote Sensing
IEEE International Conference on Image Processing (ICIP 2022). DOI: 10.1109/ICIP46576.2022.9897500
Georgii Mikriukov, Mahdyar Ravanbakhsh, Begüm Demir.
Unveiling the Anatomy of Adversarial Attacks: Concept-Based XAI Dissection of CNNs
World Conference on Explainable Artificial Intelligence (xAI 2024). DOI: 10.1007/978-3-031-63787-2_6
Georgii Mikriukov, Gesina Schwalbe, Franz Motzkus, Korinna Bade.
Revealing Similar Semantics Inside CNNs: An Interpretable Concept-Based Comparison of Feature Spaces
Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2023). DOI: 10.1007/978-3-031-74630-7_1
Georgii Mikriukov, Gesina Schwalbe, Christian Hellert, Korinna Bade.
Concept-Based Explanations in Computer Vision: Where Are We and Where Could We Go?
Explainable Computer Vision Workshop @ ECCV (eXCV 2024). DOI: 10.1007/978-3-031-92648-8_17
Jae Hee Lee, Georgii Mikriukov, Gesina Schwalbe, Stefan Wermter, Diedrich Wolter.
Uncertainty Gating for Cost-Aware Explainable Artificial Intelligence
World Conference on Explainable Artificial Intelligence (xAI 2026), accepted. DOI: 10.48550/arXiv.2603.29915
Georgii Mikriukov, Grégoire Montavon, Marina M.-C. Höhne.
On Background Bias of Post-Hoc Concept Embeddings in Computer Vision DNNs
World Conference on Explainable Artificial Intelligence (xAI 2025). DOI: 10.1007/978-3-032-08330-2_3
Gesina Schwalbe, Georgii Mikriukov, Edgar Heinert, Stavros Gerolymatos, Mert Keser, Alois Knoll, Matthias Rottmann, Annika Mütze.
Locally Testing Model Detections for Semantic Global Concepts
World Conference on Explainable Artificial Intelligence (xAI 2024). DOI: 10.1007/978-3-031-63787-2_8
Franz Motzkus, Georgii Mikriukov, Christian Hellert, Ute Schmid.
Patents and Inventions
Additional patent applications filed in 2025–2026 with Continental AG / AUMOVIO in adversarial robustness, model verification, and interpretability for automotive systems; currently under examination.
Academic Service
- Reviewer, International Journal of Computer Vision (IJCV), 2025
- Reviewer, Neural Networks (Elsevier), 2026
- Reviewer, Information Fusion (Elsevier), 2023
- Program Committee Member, KI 2026 (49th German Conference on Artificial Intelligence)
- Reviewer, Explainable Computer Vision Workshop @ ECCV (eXCV), 2025, 2026