NDSS 2026: Cutting-Edge Internet Security Research

The Network and Distributed System Security (NDSS) Symposium remains a cornerstone event for cybersecurity experts, drawing researchers from academia, industry, and government to share pioneering work. Held annually in San Diego, the 2026 edition—spanning February 23 to 27 at the Wyndham San Diego Bayside—boasts over 200 peer-reviewed papers, workshops, and keynotes. This year’s program addresses pressing challenges in an era dominated by AI-driven threats, sophisticated distributed attacks, and evolving privacy concerns. From machine learning model vulnerabilities to novel side-channel exploits, NDSS 2026 showcases research that not only identifies weaknesses but proposes practical defenses, shaping the future of secure networked systems.

Evolution of NDSS: A Decade of Innovation

Since its inception, NDSS has grown exponentially, evolving from a modest gathering into a premier venue with thousands of submissions. The 2026 symposium accepts around 20% of papers, reflecting rigorous peer review. Topics span authentication, cryptography, IoT security, blockchain, and web vulnerabilities, but emerging themes like AI security and large language model (LLM) privacy dominate. Past events, such as NDSS 2019 with 89 papers, pale in comparison to 2026’s scale, underscoring the field’s rapid expansion amid rising cyber threats.

Organized by the Internet Society and hosted at venues fostering collaboration, NDSS emphasizes real-world applicability. Attendees benefit from sessions on cutting-edge topics, poster presentations, and workshops that bridge theory and practice. This year’s focus on black-box attacks and robust detection systems highlights the symposium’s role in countering adversaries who operate with limited knowledge of defenses.

Breaking ML Defenses: The Rise of Adversarial Traffic Evasion

Machine learning-based detection has revolutionized malicious traffic identification, surpassing traditional rule-based systems in handling complex attacks. However, NDSS 2026 reveals critical robustness gaps. A standout paper introduces NetMasquerade, a reinforcement learning (RL) framework that crafts adversarial examples to mimic benign traffic, evading detectors with minimal alterations.

NetMasquerade begins with Traffic-BERT, a pre-trained model using a network-specific tokenizer and attention mechanisms to capture diverse benign patterns. Integrated into RL, it optimizes malicious packet sequences for stealth. Across 80 scenarios against six detectors—including empirically and certifiably robust ones—NetMasquerade achieves a 96.65% success rate. Both brute-force and stealthy attacks succeed, with low-latency generation proving real-world viability.

• Key Innovations: Black-box operation without model details; minimal packet changes (e.g., timing tweaks); outperforms prior attacks reliant on encrypted protocols or white-box access.
• Implications: Urges retraining ML models with adversarial examples and hybrid rule-ML approaches.

This work exposes how attackers exploit ML’s pattern-matching without needing internal knowledge, prompting a reevaluation of detection paradigms.

Combating Carpet Bombing DDoS: Server-Gateway Synergy

Distributed Denial-of-Service (DDoS) attacks, particularly carpet bombing variants, overwhelm networks by flooding multiple servers indiscriminately. Traditional detectors falter due to distributed traffic dilution. Enter NetRadar, a novel system leveraging server-gateway cooperation for precise victim identification and mitigation.

NetRadar aggregates traffic features from gateways and server-side metrics across victim networks, enabling cross-server analysis. It introduces a universal server feature set adaptable to services like web, email, and databases, paired with robust training to handle runtime mismatches. In evaluations, NetRadar excels at pinpointing victims even under heavy bombardment, reducing false positives.

By design, NetRadar scales to large networks, offering deployable defenses against evolving DDoS tactics.

Privacy Nightmares in LLMs: Inference Attacks Refined

Large language models power countless applications, but membership inference attacks (MIAs) threaten user privacy by deducing training data presence. NDSS 2026 presents ICP-MIA, a black-box MIA leveraging in-context prompting for superior accuracy.

ICP-MIA uses prompt-engineered reference data to query LLMs, exploiting output distributions for membership signals. Tested on multiple LLMs across tasks, it outperforms baselines at low false-positive rates. Analyses reveal impacts of data alignment, model architecture, PEFT setups, and training schedules on efficacy, providing a framework for auditing deployed models.

• Performance Highlights: Up to 20% higher AUC than prior black-box MIAs; effective on fine-tuned LLMs.
• Defensive Insights: Differential privacy tuning and output sanitization as countermeasures.

This research underscores the need for privacy-by-design in AI deployments.

Side-Channel Exploits: Precision Threats from Page Caches

Hardware side-channels persist as stealthy vectors. A NDSS paper dissects page cache attacks, achieving unprecedented resolution. For high-frequency signals, it detects CPUID instructions across processes with 4 kB spatial and 0.8 μs temporal granularity—six orders beyond prior art.

Low-frequency attacks target inter-keystroke timings and events, while website fingerprinting hits 90.54% F1 in top-100 closed-world tests. These feats exploit shared caches in multi-tenant environments, bypassing isolation.

1. High-Freq Detection: CPUID across processes.
2. Low-Freq: Keystroke timing (0.8 μs).
3. Fingerprinting: 90.54% F1 score.

Mitigations like cache partitioning and flushing are proposed, but trade-offs in performance demand balanced implementations.

Deobfuscation Breakthroughs: Unmasking JavaScript Malware

Malicious JavaScript often hides via obfuscation. JSIMPLIFIER tackles this with a dataset of 44,421 samples (23,212 malicious, 21,209 benign). It processes 100% of 20 obfuscation techniques, reduces complexity by 88.2%, and boosts readability 4x, validated by LLMs.

Released publicly, this tool aids dynamic analysis and signature generation.

Future Directions and Symposium Impact

NDSS 2026’s papers signal a maturing field: AI defenses must anticipate evasion, DDoS tools need cross-layer intel, LLMs require privacy audits, and side-channels demand hardware-software fixes. With CyLab/CMU contributions and broad participation, the event drives deployable solutions.

Frequently Asked Questions

What is NDSS Symposium?

NDSS is a top-tier conference on network and distributed system security, featuring peer-reviewed papers and workshops.

When and where is NDSS 2026?

February 23–27, 2026, at Wyndham San Diego Bayside, California.

How many papers are presented?

Over 200, covering diverse security topics.

Are proceedings publicly available?

Yes, via the official NDSS website post-event.

Who should attend NDSS?

Researchers, practitioners, and policymakers in cybersecurity.