We've introduced a new feature on FHERMA.io that we believe will be useful for researchers and developers: Kernel Radars.
Kernel Radars allow you to track developments related to a specific kernel, including new research, algorithms, implementations, optimizations, security attacks, and other relevant findings.
Each radar continuously monitors multiple sources, identifies significant developments, and connects them to the corresponding kernel.
The goal is to provide a continuously updated view of the research and technological landscape surrounding each kernel, making it easier to discover relevant advances, understand their significance, and assess their potential impact on performance, security, and implementation.
Radars powered by monstream.io .
Explore an example - CKKS Bootstrapping Radar:
fherma.io/kernels/ckks-boots…
NVIDIA × Fair Math Challenges are LIVE on FHERMA!
We're excited to collaborate with @NVIDIAHPCDev to bring GPU acceleration layer support to the FHE ecosystem using NVIDIA's cuPQC.
To kickstart this initiative, we're launching two new developer challenges with GPU compute:
- Polynomial Multiplication: Build high-performance polynomial multiplication on cuPQC's bigint backend. (Great entry point for GPU/CUDA devs!)
- Key Switching in CKKS: Optimize one of FHE's biggest bottlenecks using GPU-native arithmetic.
Whether you're an FHE researcher or a CUDA/GPU dev, you can help build open-source infrastructure that projects worldwide will use.
Join the challenges: fherma.io
#FHE #Cryptography #CUDA #GPU #cuPQC #NVIDIA #OpenSource
Join us next week for the OpenFHE webinar on Efficient Large-Integer Arithmetic for FHE.
We’ll present recent results from the collaboration between NVIDIA, Fair Math and Duality Technologies, including cuPQC, FHERMA, and our recent joint work on large-integer arithmetic.
Register here: brighttalk.com/webcast/19414…
1/8 Today, we’re introducing a completely new FHERMA.
What started as a challenge and benchmarking platform for FHE is evolving into a much bigger infrastructure layer: a common system for discovering, evaluating and using the state-of-the-art building blocks/components behind modern privacy technologies and the applications built on top of them.
Happy to share our collaboration with @nvidia and @DualityTech research teams: "Efficient Large-Integer Arithmetic for FHE".
Efficient large-integer arithmetic is a key building block of modern fully homomorphic encryption (FHE). In this paper, we discuss the algorithms and implementation techniques behind high-performance arithmetic for vectorized FHE schemes, examine the trade-offs between different approaches, and highlight promising directions for future improvements.
It was a pleasure collaborating with researchers from NVIDIA and Duality Technologies on this work.
1/ We’re opening private testing for Manaslu: a privacy-preserving data retrieval framework for AI agents.
AI agents today continuously read from public data sources: markets, blockchains, APIs, search systems, datasets, knowledge bases.
Every retrieval request leaks sensitive information about the agent:
- intent
- strategy
- objectives
- behavioral patterns
That’s the problem Manaslu is designed to solve.
We are launching a major new competition on FHERMA focused on privacy-preserving AI for research groups and individual researchers.
If you are interested in exploring FHE for AI applications and would like to participate, please complete the form (forms.gle/stD41f8fWYRm3G5EA) by October 15.
Pre-registration is required to participate in the competition.
Some updates regarding our Palitra.ai experiment on how AI can keep secrets:
A @MistralAI-based agent struggled to follow system instructions — users extracted the secret 50+ times last week. A comparable @OpenAI-based agent proved far more resilient.
@grok-based agent performed better than the @deepseek_ai-based agent in keeping secrets.
More detailed analysis & trends: huggingface.co/blog/vln21/pa…
It is essential to develop methods and technologies as soon as possible that will allow AI to autonomously work with sensitive/secret data.
