🕵️ RESEARCH & DEEP DIVES
FLINT fingerprints federated learning models via 5G PHY-layer side channels — arxiv.org
Researchers developed FLINT to identify federated learning (FL) model types using 5G physical layer metadata.- Targets federated learning (FL) models running over 5G cellular networks
- Exploits PHY-layer scheduling metadata broadcast on the Physical Downlink Control Channel (PDCCH)
- Infers model architecture families like CNNs, RNNs, and Transformers without network-layer visibility
- Uses decoding of PDCCH, mapping changing RNTIs to devices, and multi-view temporal modeling
- Demonstrated on srsRAN-based 5G testbed with a macro F1-score of 0.930 for classification
KASS framework enables executable attack synthesis and simulation for smart contracts — arxiv.org
Researchers developed KASS, a framework that generates and simulates executable smart contract exploits.- Applies to smart contracts across multiple vulnerability categories
- Generates executable exploits by decomposing attack synthesis into planning, generation, and testing
- Uses retrieval-augmented planning with real-world audit knowledge for attack strategy
- Validates exploits with formal constraints and dual-loop refinement to fix errors and replan
- Successfully exploits 94.23% of tested contracts and validates 9 of 11 real-world CVE cases
Researchers demonstrate natural backdoor attacks on speech recognition models — arxiv.org
Natural sounds can trigger stealthy backdoor attacks on speech recognition systems.- Targets speech recognition models trained on two datasets and three architectures
- Uses ordinary natural or daily life sounds as stealthy backdoor triggers
- Achieves near 100% attack success with only 5% poisoned training samples
- Triggers activate automatically by corresponding natural sounds without degrading normal performance
- Attack effectiveness influenced by poisoning rate, trigger duration, and blend ratio
dicomhawk-medical-imaging-cyber-deception — arxiv.org
dicom-internet-exposed-services-measurement — arxiv.org
dos-quality-attack-5g-nr-side-channels — arxiv.org
llm-chain-of-thought-jailbreak-research — arxiv.org
black-box-visual-attacks-multimodal-ai — arxiv.org
prompt-syntax-secure-code-generation-llms — arxiv.org
code-poisoning-property-inference-attacks — arxiv.org
open-weight-llms-threat-info-autonomous-vehicles — arxiv.org