2025 BrainChip Discussion, page-4290

This detailed paper is 29 pages, so here's an explanation summary for the time poor and layman:

*Introduction

The research paper introduces NQAL (Neuromorphic Quantum Adversarial Learning), a cutting-edge defense framework developed to detect hidden threats within DNS over HTTPS (DoH) traffic. While DoH enhances privacy by encrypting DNS queries, it also makes it more difficult for conventional cybersecurity systems to identify malicious activities.

To address the challenges of encrypted DNS traffic, the NQAL model combines Spiking Neural Networks (SNNs) with Quantum Noise Injection (QNI) for real-time anomaly detection. Implemented on neuromorphic chips including BrainChip’s Akida and Intel’s Loihi 2, the system achieved best-ever performance with Akida outperforming all platforms — with the Akida NQAL model delivering a 99.32% accuracy rate that surpassed CNNs, Transformers, and Quantum HNNs. It also showed strong resistance to adversarial attacks (FGSM, PGD) and used 12× less energy than traditional GPU-based models, making it a powerful and efficient solution for real-time DoH threat detection.

In simpler terms:

• "Quantum noise" refers to the unpredictable, natural randomness that occurs at the atomic or subatomic level (like particles behaving differently every time you measure them).

• "Injection" means intentionally adding this kind of randomness into the system.

Why inject noise at all?

Adding carefully designed noise can actually help machine learning models by:

• Preventing them from becoming too rigid or overfitted to specific training data.

• Making it harder for adversarial attackers to trick the model.

• Helping the system explore more possibilities and better generalize to unseen data.

And in the context of Akida and NQAL?

In NQAL:

• QNI helps Akida learn to resist adversarial attacks, such as when hackers try to subtly alter traffic to fool detection systems.

• This combination gives Akida the ability to:

  • Detect brand-new, never-before-seen attacks (known as zero-day threats).

  • Withstand attacks from future quantum computers (post-quantum resilience).

Traditional security models often rely on recognizing known patterns, which leaves them vulnerable to new or highly advanced attacks. In contrast, QNI enables Akida to "think outside the box." It doesn't just look for familiar threats — it can also identify unusual or suspicious behavior that hasn’t been encountered before. This makes Akida more resilient against both surprise attacks and the highly complex threats expected in a post-quantum future.

Analogy:

Think of QNI like teaching a dog to recognize a person not just in a clean photo, but also in the rain, in fog, with a hat, or after they’ve shaved their beard. You’re training the system to handle real-world messiness and deception — and QNI provides that simulated unpredictability. Akida is the dog that learns what the person really looks like, even if their appearance changes or the environment is messy. It doesn’t just memorize patterns — it understands what to look for, making it much harder to fool.

Akida's Role in NQAL

Core functions Akida performs in the system:

• Executes Spiking Neural Networks with Dynamic Spiking Graph Attention (DSGAT): Focuses on the most important data patterns in real time, similar to how the brain tunes out distractions.

• Uses Spike-Timing-Dependent Plasticity (STDP): Learns from the timing of events, allowing it to adapt quickly to new patterns.

• Processes encrypted DNS traffic efficiently: Delivers high speed and accuracy with very low power use.

Performance Evaluation

Akida was benchmarked against a traditional GPU (NVIDIA V100), TPUv4, and Intel’s Loihi 2 neuromorphic chip. The performance results are summarized below:

	Metric	GPU (V100)	TPUv4	Loihi 2	Akida
1	Accuracy	89.2%	91.5%	98.7%	99.1%
2	Latency	3.1 ms	2.8 ms	0.9 ms	0.7 ms
3	Power Use	45 W	32 W	4 W	3 W
4	Throughput	1,200 QPS	1,500 QPS	9,800 QPS	12,400 QPS

Akida outperformed all other platforms in every major category, offering the best accuracy, lowest latency, lowest power consumption, and highest throughput.

Integration into NQAL Architecture

Akida detects advanced threat types including:

• DNS Tunneling - a sneaky way that hackers hide data within regular web requests, bypassing firewalls

• Command-and-Control (C2) Beaconing - when infected devices regularly connect to a hacker's server to receive instructions

• Domain Generation Algorithms (DGAs) - used by malware to create constantly changing domain names, making them harder to track or block

Akida detects these threats by identifying unusual timing patterns in encrypted traffic, a task that traditional systems often miss. It does so with remarkable speed, accuracy, and energy efficiency. Its DSGAT layer dynamically focuses on important patterns in this data, and the STDP mechanism allows the model to learn and adapt over time.

Threat Robustness

Akida’s hardware-level integration with Quantum Noise Injection (QNI) and STDP learning enables the system to defend against a wide variety of sophisticated cyber threats.

	Threat Type	Detection / Robustness
1	DNS Tunneling	98.7%
2	C2 Beaconing	99.2%
3	Adversarial Attacks	94%
4	Zero-day DGA	97.5%
5	Quantum Perturbation	100% blocked

Akida demonstrated exceptional threat detection and resilience, thanks to its integration of Quantum Noise Injection (QNI) and Spike-Timing-Dependent Plasticity (STDP). It consistently detected a wide range of advanced cyber threats with high accuracy.

Energy and Cost Benefits

Compared to traditional legacy systems, deploying NQAL on Akida leads to significant improvements in cost-efficiency, false positives, and system security.

	Metric	Legacy Systems	NQAL on Akida	Improvement
1	False Positives / Day	420	18	23.3× fewer
2	Energy Cost / Month	$2,800	$240	11.7× lower
3	Attack Surface Exposure	8.7 vectors	0.2 vectors	43.5× safer
4	Compliance Coverage	62%	98%	+36% improvement

What Version of Akida was Used:

The research paper does not explicitly state the version of Akida used. However, based on the performance metrics and the fact that it compares Akida directly to Intel Loihi 2, it is highly likely that the version used was from the Akida 2nd Generation (Akida 2.0) family — most likely Akida-P or Akida-S, given the high throughput (12,400 QPS) and low latency (0.7 ms).
Supporting reasoning:

• The paper emphasizes post-quantum resilience, spike-based attention (DSGAT), and QNI integration, all of which align with the newer architectural features of Akida 2.0.

• Akida 2.0 supports higher-performance SNN tasks and more advanced neural processing units (NPUs), which are required to achieve the reported results.

While not confirmed in the paper, the evidence strongly suggests the use of Akida 2.0, most likely the Akida-P (Performance) variant.

Final Takeaways

BrainChip’s Akida chip demonstrates best-in-class performance for detecting threats in encrypted internet traffic. It surpasses both traditional and neuromorphic competitors in every measured category. Its integration with bio-inspired learning and quantum-resistant defenses positions it as a foundational technology for future cybersecurity systems, particularly those requiring real-time response and low energy consumption at the edge.
*gpt4o