Latest news with #LoRA
The Sun
02-07-2025
- Business
- The Sun
CyCraft Launches XecGuard: LLM Firewall for Trustworthy AI
TAIPEI, TAIWAN - Media OutReach Newswire - 1 July 2025 - CyCraft, a leading AI cybersecurity firm, today announced the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend Large Language Models (LLMs). XecGuard's introduction marks a pivotal moment for secure, trustworthy AI, addressing the critical security challenges posed by the rapid adoption of LLMs. Trustworthy AI Matters The transformative power of Large Language Models (LLMs) brings significant security uncertainty, requiring enterprises to urgently safeguard their AI models from malicious attacks like prompt injection, prompt extraction, and jailbreak attempts. Historically, AI security has been an 'optional add-on' rather than a fundamental feature, leaving valuable AI and data exposed. This oversight can compromise sensitive data, undermine service stability, and erode customer trust. CyCraft emphasizes that 'AI security must be a standard feature—not an optional add-on,' believing it's paramount for delivering stable and trustworthy intelligent services. The Imminent Need for Proactive AI Defense The need for immediate and effective AI security is more critical than ever before. As AI becomes increasingly embedded in core business operations, the attack surface expands exponentially, making proactive defenses an absolute necessity. CyCraft has leveraged its extensive 'battle-tested expertise across critical domains—including government, finance, and high-tech manufacturing' to precisely address these emerging AI-specific threats. The development of XecGuard signifies a shift from 'using AI to tackle cybersecurity challenges' to now 'using AI to protect AI' , ensuring that security and resilience are embedded from day one. 'AI security must be a standard feature—not an optional add-on,' stated Benson Wu, CEO, highlighting XecGuard's resilience and integration of experience from defending critical sectors. Jeremy Chiu, CTO and Co-Founder, emphasized, 'In the past, we used AI to tackle cybersecurity challenges; now, we're using AI to protect AI,' adding that XecGuard enables enterprises to confidently adopt AI and deliver trustworthy services. PK Tsung, CISO, concluded, 'With XecGuard, we're empowering enterprises to embed security and resilience from day one' as part of their vision for the world's most advanced AI security platform. CyCraft's Solution: XecGuard Empowers Secure AI Deployment CyCraft leads with the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend LLMs. XecGuard provides robust protection against prompt injection, prompt extraction, and jailbreak attacks, ensuring enterprise-grade resilience for AI models. Its seamless deployment allows instant integration with any LLM without architectural modification, delivering powerful autonomous defense out of the box. XecGuard is available as a SaaS, an OpenAI-compatible LLM firewall on your cloud (e.g., AWS or Cloudflare Workers AI), or an embedded firewall for on-premises, NVIDIA-powered custom LLM servers. Rigorously validated on major open-source models like Llama 3B, Qwen3 4B, Gemma3 4B, and DeepSeek 8B, it consistently improves security resilience while preserving core performance, enabling even small models to achieve protection comparable to large commercial-grade systems. Real-world validation through collaboration with APMIC, an NVIDIA partner, integrated XecGuard into the F1 open-source model, demonstrating an average 17.3% improvement in overall security defense scores and up to 30.1% in specific attack scenarios via LLM Red Teaming exercises. With XecGuard and the Safety LLM service, CyCraft delivers enterprise-grade AI security, accelerating the adoption of resilient and trustworthy AI across industries, empowering organizations to deploy AI securely, protect sensitive data, and drive innovation with confidence. Even small models gain enterprise-level defenses, approaching large commercial-grade performance.
Arabian Post
01-07-2025
- Business
- Arabian Post
CyCraft Launches XecGuard: LLM Firewall for Trustworthy AI
CyCraft Co-Founders (from left to right): Benson Wu (CEO), Jeremy Chiu (CTO), and PK Tsung (CISO) are leading the mission to build the world's most advanced AI security platform. TAIPEI, TAIWAN – Media OutReach Newswire – 1 July 2025 – CyCraft, a leading AI cybersecurity firm, today announced the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend Large Language Models (LLMs). XecGuard's introduction marks a pivotal moment for secure, trustworthy AI, addressing the critical security challenges posed by the rapid adoption of LLMs. Trustworthy AI Matters The transformative power of Large Language Models (LLMs) brings significant security uncertainty, requiring enterprises to urgently safeguard their AI models from malicious attacks like prompt injection, prompt extraction, and jailbreak attempts. Historically, AI security has been an 'optional add-on' rather than a fundamental feature, leaving valuable AI and data exposed. This oversight can compromise sensitive data, undermine service stability, and erode customer trust. CyCraft emphasizes that 'AI security must be a standard feature—not an optional add-on,' believing it's paramount for delivering stable and trustworthy intelligent services. The Imminent Need for Proactive AI Defense The need for immediate and effective AI security is more critical than ever before. As AI becomes increasingly embedded in core business operations, the attack surface expands exponentially, making proactive defenses an absolute necessity. CyCraft has leveraged its extensive 'battle-tested expertise across critical domains—including government, finance, and high-tech manufacturing' to precisely address these emerging AI-specific threats. The development of XecGuard signifies a shift from 'using AI to tackle cybersecurity challenges' to now 'using AI to protect AI' , ensuring that security and resilience are embedded from day one. ADVERTISEMENT 'AI security must be a standard feature—not an optional add-on,' stated Benson Wu, CEO, highlighting XecGuard's resilience and integration of experience from defending critical sectors. Jeremy Chiu, CTO and Co-Founder, emphasized, 'In the past, we used AI to tackle cybersecurity challenges; now, we're using AI to protect AI,' adding that XecGuard enables enterprises to confidently adopt AI and deliver trustworthy services. PK Tsung, CISO, concluded, 'With XecGuard, we're empowering enterprises to embed security and resilience from day one' as part of their vision for the world's most advanced AI security platform. CyCraft's Solution: XecGuard Empowers Secure AI Deployment CyCraft leads with the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend LLMs. XecGuard provides robust protection against prompt injection, prompt extraction, and jailbreak attacks, ensuring enterprise-grade resilience for AI models. Its seamless deployment allows instant integration with any LLM without architectural modification, delivering powerful autonomous defense out of the box. XecGuard is available as a SaaS, an OpenAI-compatible LLM firewall on your cloud (e.g., AWS or Cloudflare Workers AI), or an embedded firewall for on-premises, NVIDIA-powered custom LLM servers. Rigorously validated on major open-source models like Llama 3B, Qwen3 4B, Gemma3 4B, and DeepSeek 8B, it consistently improves security resilience while preserving core performance, enabling even small models to achieve protection comparable to large commercial-grade systems. Even small models gain enterprise-level defenses, approaching large commercial-grade performance. Real-world validation through collaboration with APMIC, an NVIDIA partner, integrated XecGuard into the F1 open-source model, demonstrating an average 17.3% improvement in overall security defense scores and up to 30.1% in specific attack scenarios via LLM Red Teaming exercises. With XecGuard and the Safety LLM service, CyCraft delivers enterprise-grade AI security, accelerating the adoption of resilient and trustworthy AI across industries, empowering organizations to deploy AI securely, protect sensitive data, and drive innovation with confidence. To learn more about how XecGuard can protect your LLMs and to request a demo, visit: Hashtag: #CyCraft #LLMFirewall #AISecurity The issuer is solely responsible for the content of this announcement. About CyCraft Technology CyCraft is a leading AI-driven cybersecurity company in the Asia-Pacific region. Trusted by hundreds of organizations in defense, finance, and semiconductor industries, our AI is designed to prevent, preempt, and protect against cyber threats. Our expertise has been recognized by top-tier institutions like Gartner and IDC and showcased at prestigious global conferences, including Black Hat, DEFCON, EMNLP, and Code Blue.
Associated Press
01-07-2025
- Business
- Associated Press
CyCraft Launches XecGuard: LLM Firewall for Trustworthy AI
TAIPEI, TAIWAN - Media OutReach Newswire - 1 July 2025 - CyCraft, a leading AI cybersecurity firm, today announced the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend Large Language Models (LLMs). XecGuard's introduction marks a pivotal moment for secure, trustworthy AI, addressing the critical security challenges posed by the rapid adoption of LLMs. CyCraft Co-Founders (from left to right): Benson Wu (CEO), Jeremy Chiu (CTO), and PK Tsung (CISO) are leading the mission to build the world's most advanced AI security platform. Trustworthy AI Matters The transformative power of Large Language Models (LLMs) brings significant security uncertainty, requiring enterprises to urgently safeguard their AI models from malicious attacks like prompt injection, prompt extraction, and jailbreak attempts. Historically, AI security has been an 'optional add-on' rather than a fundamental feature, leaving valuable AI and data exposed. This oversight can compromise sensitive data, undermine service stability, and erode customer trust. CyCraft emphasizes that 'AI security must be a standard feature—not an optional add-on,' believing it's paramount for delivering stable and trustworthy intelligent services. The Imminent Need for Proactive AI Defense The need for immediate and effective AI security is more critical than ever before. As AI becomes increasingly embedded in core business operations, the attack surface expands exponentially, making proactive defenses an absolute necessity. CyCraft has leveraged its extensive 'battle-tested expertise across critical domains—including government, finance, and high-tech manufacturing' to precisely address these emerging AI-specific threats. The development of XecGuard signifies a shift from 'using AI to tackle cybersecurity challenges' to now 'using AI to protect AI' , ensuring that security and resilience are embedded from day one. 'AI security must be a standard feature—not an optional add-on,' stated Benson Wu, CEO, highlighting XecGuard's resilience and integration of experience from defending critical sectors. Jeremy Chiu, CTO and Co-Founder, emphasized, 'In the past, we used AI to tackle cybersecurity challenges; now, we're using AI to protect AI,' adding that XecGuard enables enterprises to confidently adopt AI and deliver trustworthy services. PK Tsung, CISO, concluded, 'With XecGuard, we're empowering enterprises to embed security and resilience from day one' as part of their vision for the world's most advanced AI security platform. CyCraft's Solution: XecGuard Empowers Secure AI Deployment CyCraft leads with the global launch of XecGuard, the industry's first plug-and-play LoRA security module purpose-built to defend LLMs. XecGuard provides robust protection against prompt injection, prompt extraction, and jailbreak attacks, ensuring enterprise-grade resilience for AI models. Its seamless deployment allows instant integration with any LLM without architectural modification, delivering powerful autonomous defense out of the box. XecGuard is available as a SaaS, an OpenAI-compatible LLM firewall on your cloud (e.g., AWS or Cloudflare Workers AI), or an embedded firewall for on-premises, NVIDIA-powered custom LLM servers. Rigorously validated on major open-source models like Llama 3B, Qwen3 4B, Gemma3 4B, and DeepSeek 8B, it consistently improves security resilience while preserving core performance, enabling even small models to achieve protection comparable to large commercial-grade systems. Even small models gain enterprise-level defenses, approaching large commercial-grade performance. Real-world validation through collaboration with APMIC, an NVIDIA partner, integrated XecGuard into the F1 open-source model, demonstrating an average 17.3% improvement in overall security defense scores and up to 30.1% in specific attack scenarios via LLM Red Teaming exercises. With XecGuard and the Safety LLM service, CyCraft delivers enterprise-grade AI security, accelerating the adoption of resilient and trustworthy AI across industries, empowering organizations to deploy AI securely, protect sensitive data, and drive innovation with confidence. To learn more about how XecGuard can protect your LLMs and to request a demo, visit: Hashtag: #CyCraft #LLMFirewall #AISecurity The issuer is solely responsible for the content of this announcement. About CyCraft Technology CyCraftis a leading AI-driven cybersecurity company in the Asia-Pacific region. Trusted by hundreds of organizations in defense, finance, and semiconductor industries, our AI is designed to prevent, preempt, and protect against cyber threats. Our expertise has been recognized by top-tier institutions like Gartner and IDC and showcased at prestigious global conferences, including Black Hat, DEFCON, EMNLP, and Code Blue.
Geeky Gadgets
16-06-2025
- Geeky Gadgets
How to Fine Tune your own LLM using LoRA (on a Custom dataset)
Imagine unlocking the full potential of a massive language model, tailoring it to your unique needs without breaking the bank or requiring a supercomputer. Sounds impossible? It's not. Thanks to Low-Rank Adaptation (LoRA), fine-tuning large language models (LLMs) has become more accessible than ever. Whether you're a developer aiming to build a hyper-specific chatbot or a researcher looking to extract insights from niche datasets, LoRA offers a streamlined, resource-efficient way to customize LLMs. Gone are the days of needing vast computational power to adapt these models—LoRA's innovative approach lets you focus on creativity and precision, not hardware limitations. Nicholas Renotte walks you through the process of fine-tuning your own LLM using LoRA on a custom dataset. You'll discover how to prepare your data, set up an efficient training environment, and integrate LoRA's modular layers to achieve task-specific results—all while preserving the original model's versatility. Along the way, you'll learn why LoRA is transforming how we approach fine-tuning, offering faster training times and reduced hardware demands. By the end, you'll not only understand the mechanics of LoRA but also gain the confidence to apply it to your own projects. What could your fine-tuned LLM achieve? Let's explore the possibilities. Fine-Tuning LLMs with LoRA Why Choose Low-Rank Adaptation (LoRA)? LoRA is an innovative technique designed to reduce the computational and memory demands of fine-tuning large-scale models. Instead of modifying all the parameters of an LLM, LoRA introduces trainable low-rank matrices into the model's architecture. This approach enables efficient adaptation for specific tasks while preserving the model's general capabilities. The key benefits of LoRA include: Reduced hardware requirements: LoRA significantly lowers the computational burden, making fine-tuning feasible even on systems with limited resources. LoRA significantly lowers the computational burden, making fine-tuning feasible even on systems with limited resources. Faster training times: Compared to traditional fine-tuning methods, LoRA accelerates the process, saving time and effort. Compared to traditional fine-tuning methods, LoRA accelerates the process, saving time and effort. Preservation of general knowledge: The original model retains its broad capabilities, making sure versatility across multiple tasks. These advantages make LoRA an ideal choice for researchers and developers aiming to fine-tune LLMs efficiently. Preparing Your Custom Dataset The success of fine-tuning largely depends on the quality and relevance of your custom dataset. To ensure your dataset is effective: Focus on relevance: Select data that is directly aligned with the task you aim to solve. The dataset should accurately represent the problem domain. Select data that is directly aligned with the task you aim to solve. The dataset should accurately represent the problem domain. Clean and preprocess: Remove inconsistencies, duplicates, and irrelevant entries to enhance data quality and reliability. Remove inconsistencies, duplicates, and irrelevant entries to enhance data quality and reliability. Format appropriately: Structure the dataset to match the input-output format expected by the pre-trained model. This ensures seamless integration during training. For instance, if you are fine-tuning an LLM for sentiment analysis, your dataset should include labeled text samples categorized as positive, negative, or neutral. A well-prepared dataset lays the foundation for effective fine-tuning and improved model performance. Fine Tune Your Own AI using LoRA Watch this video on YouTube. Advance your skills in Large Language Models (LLMs) by reading more of our detailed content. Setting Up Your Environment Creating the right environment is essential for implementing LoRA successfully. Follow these steps to set up your environment: Select a pre-trained model: Choose an LLM that aligns with your task requirements, such as GPT-based models, BERT, or T5. Choose an LLM that aligns with your task requirements, such as GPT-based models, BERT, or T5. Install necessary frameworks: Use machine learning libraries like PyTorch or TensorFlow, making sure they support LoRA integration and provide the required tools. Use machine learning libraries like PyTorch or TensorFlow, making sure they support LoRA integration and provide the required tools. Verify computational resources: Confirm that your hardware, such as GPUs or TPUs, meets the minimum requirements for the chosen model and task. By establishing a robust environment, you can streamline the fine-tuning process and minimize potential technical challenges. Fine-Tuning with LoRA The fine-tuning process using LoRA involves several critical steps that ensure efficiency and accuracy: Integrate LoRA: Add LoRA layers to specific components of the pre-trained model, such as attention mechanisms, to enable task-specific adaptation. Add LoRA layers to specific components of the pre-trained model, such as attention mechanisms, to enable task-specific adaptation. Freeze original parameters: Keep the base model's parameters fixed to retain its general knowledge and prevent unnecessary modifications. Keep the base model's parameters fixed to retain its general knowledge and prevent unnecessary modifications. Train on your dataset: Use the prepared dataset to train the LoRA parameters. Monitor the training process closely to avoid overfitting and ensure steady progress. Use the prepared dataset to train the LoRA parameters. Monitor the training process closely to avoid overfitting and ensure steady progress. Validate the model: Test the fine-tuned model on a validation set to evaluate its performance and identify areas for improvement. LoRA's modular design allows you to fine-tune multiple tasks on the same base model by swapping out the low-rank matrices. This flexibility makes it a cost-effective and reusable solution for various applications. Optimizing the Fine-Tuning Process To achieve the best results, it is essential to optimize the fine-tuning process. Consider the following strategies: Experiment with hyperparameters: Adjust learning rates, batch sizes, and other settings to identify the optimal configuration for your task. Adjust learning rates, batch sizes, and other settings to identify the optimal configuration for your task. Use early stopping: Halt training when performance metrics plateau to prevent overfitting and save computational resources. Halt training when performance metrics plateau to prevent overfitting and save computational resources. Monitor key metrics: Track metrics such as accuracy, precision, recall, or task-specific measures to evaluate progress and make data-driven adjustments. These optimization techniques ensure that your fine-tuned model generalizes well to unseen data while maintaining high performance on the target task. Evaluating Your Fine-Tuned Model Evaluation is a crucial step to verify that your fine-tuned LLM meets the desired objectives. Use a test dataset that reflects real-world scenarios to assess the model's performance. Depending on the task, consider the following metrics: F1-score: A balanced measure of precision and recall, particularly useful for classification tasks. A balanced measure of precision and recall, particularly useful for classification tasks. BLEU: Evaluates the quality of generated text in tasks such as translation by comparing it to reference outputs. Evaluates the quality of generated text in tasks such as translation by comparing it to reference outputs. ROUGE: Measures the quality of text summarization by comparing generated summaries to reference texts. Additionally, compare the fine-tuned model's performance with the baseline results of the pre-trained model. This comparison helps quantify improvements and highlights the effectiveness of the fine-tuning process. Applications of Fine-Tuned LLMs Fine-tuned LLMs offer a wide range of applications across various industries, allowing tailored solutions for specific challenges. Some practical use cases include: Customer Support: Develop intelligent chatbots capable of providing accurate and context-aware responses to user queries. Develop intelligent chatbots capable of providing accurate and context-aware responses to user queries. Content Generation: Automate the creation of customized content for marketing, education, or entertainment purposes. Automate the creation of customized content for marketing, education, or entertainment purposes. Medical Research: Summarize complex medical literature to assist healthcare professionals in making informed decisions. Summarize complex medical literature to assist healthcare professionals in making informed decisions. Sentiment Analysis: Analyze public opinion on products, services, or events using social media or survey data. By fine-tuning LLMs, organizations can address specific needs, enhance efficiency, and deliver innovative solutions tailored to their objectives. Media Credit: Nicholas Renotte Filed Under: AI, Guides Latest Geeky Gadgets Deals Disclosure: Some of our articles include affiliate links. If you buy something through one of these links, Geeky Gadgets may earn an affiliate commission. Learn about our Disclosure Policy.
The Sun
11-06-2025
- Science
- The Sun
PolyU develops novel multi-modal agent to facilitate long video understanding by AI, accelerating development of generative AI-assisted video analysis
HONG KONG SAR - Media OutReach Newswire - 10 June 2025 - While Artificial Intelligence (AI) technology is evolving rapidly, AI models still struggle with understanding long videos. A research team from The Hong Kong Polytechnic University (PolyU) has developed a novel video-language agent, VideoMind, that enables AI models to perform long video reasoning and question-answering tasks by emulating humans' way of thinking. The VideoMind framework incorporates an innovative Chain-of-Low-Rank Adaptation (LoRA) strategy to reduce the demand for computational resources and power, advancing the application of generative AI in video analysis. The findings have been submitted to the world-leading AI conferences. Videos, especially those longer than 15 minutes, carry information that unfolds over time, such as the sequence of events, causality, coherence and scene transitions. To understand the video content, AI models therefore need not only to identify the objects present, but also take into account how they change throughout the video. As visuals in videos occupy a large number of tokens, video understanding requires vast amounts of computing capacity and memory, making it difficult for AI models to process long videos. Prof. Changwen CHEN, Interim Dean of the PolyU Faculty of Computer and Mathematical Sciences and Chair Professor of Visual Computing, and his team have achieved a breakthrough in research on long video reasoning by AI. In designing VideoMind, they made reference to a human-like process of video understanding, and introduced a role-based workflow. The four roles included in the framework are: the Planner, to coordinate all other roles for each query; the Grounder, to localise and retrieve relevant moments; the Verifier, to validate the information accuracy of the retrieved moments and select the most reliable one; and the Answerer, to generate the query-aware answer. This progressive approach to video understanding helps address the challenge of temporal-grounded reasoning that most AI models face. Another core innovation of the VideoMind framework lies in its adoption of a Chain-of-LoRA strategy. LoRA is a finetuning technique emerged in recent years. It adapts AI models for specific uses without performing full-parameter retraining. The innovative chain-of-LoRA strategy pioneered by the team involves applying four lightweight LoRA adapters in a unified model, each of which is designed for calling a specific role. With this strategy, the model can dynamically activate role-specific LoRA adapters during inference via self-calling to seamlessly switch among these roles, eliminating the need and cost of deploying multiple models while enhancing the efficiency and flexibility of the single model. VideoMind is open source on GitHub and Huggingface. Details of the experiments conducted to evaluate its effectiveness in temporal-grounded video understanding across 14 diverse benchmarks are also available. Comparing VideoMind with some state-of-the-art AI models, including GPT-4o and Gemini 1.5 Pro, the researchers found that the grounding accuracy of VideoMind outperformed all competitors in challenging tasks involving videos with an average duration of 27 minutes. Notably, the team included two versions of VideoMind in the experiments: one with a smaller, 2 billion (2B) parameter model, and another with a bigger, 7 billion (7B) parameter model. The results showed that, even at the 2B size, VideoMind still yielded performance comparable with many of the other 7B size models. Prof. Chen said, 'Humans switch among different thinking modes when understanding videos: breaking down tasks, identifying relevant moments, revisiting these to confirm details and synthesising their observations into coherent answers. The process is very efficient with the human brain using only about 25 watts of power, which is about a million times lower than that of a supercomputer with equivalent computing power. Inspired by this, we designed the role-based workflow that allows AI to understand videos like human, while leveraging the chain-of-LoRA strategy to minimise the need for computing power and memory in this process.' AI is at the core of global technological development. The advancement of AI models is however constrained by insufficient computing power and excessive power consumption. Built upon a unified, open-source model Qwen2-VL and augmented with additional optimisation tools, the VideoMind framework has lowered the technological cost and the threshold for deployment, offering a feasible solution to the bottleneck of reducing power consumption in AI models. Prof. Chen added, 'VideoMind not only overcomes the performance limitations of AI models in video processing, but also serves as a modular, scalable and interpretable multimodal reasoning framework. We envision that it will expand the application of generative AI to various areas, such as intelligent surveillance, sports and entertainment video analysis, video search engines and more.'
