• AI Breakfast
  • Posts
  • How AI Can Steal Your Password With a Microphone

How AI Can Steal Your Password With a Microphone

Author

AI Breakfast
August 07, 2023

Good morning. It’s Monday, August 7th.

Did you know: The term "artificial intelligence" was first coined by John McCarthy in 1956?

In today’s email:

  • AI and Cybersecurity

  • AI in Asia

  • AI in Science and Health

  • AI in Military and Defense

  • AI Hardware

  • 5 New AI Tools

  • Latest AI Research Papers

You read. We listen. Let us know what you think of this edition by replying to this email, or DM us on Twitter.

Today’s edition is brought to you by:

Tired of explaining the same thing over and over again to your colleagues?

It’s time to delegate that work to AI.

Guidde is a GPT-powered tool that helps you explain the most complex tasks in seconds with AI generated documentation.

Turn boring documentation into stunning visual guides

  • Save valuable time by creating video documentation 11x faster

  • Share or embed your guide anywhere for your team to see

  • Simply click capture on our browser extension and the app will automatically generate step-by-step video guides complete with visuals, voiceover and call to actions.

The best part? The Guidde extension is 100% free. 

Today’s trending AI news stories

AI and Cybersecurity

This AI model can steal your data by listening to your keystrokes with 95% accuracy: Researchers from British universities have developed an AI capable of predicting users’ keystrokes with 95% accuracy by listening to the sounds produced when typing on a keyboard. The deep learning model records minute differences in keystrokes and uses waveform analysis to decipher the typed content, potentially compromising sensitive data like passwords and classified information. The study highlights the vulnerability of data, urging users to be vigilant about cybersecurity beyond avoiding phishing attacks. Hackers may no longer need a key logger to steal your password… just a microphone.

Deepfake speech is getting too good, fooling a quarter of listeners in study: A recent study by University College London highlights the growing challenge of detecting deepfake speech. The research found that humans can only identify deepfake speech with 73% accuracy, regardless of the language being used. The study warns that as deepfake technology improves, detection will become even harder, raising concerns about potential misuse in phone scams and misinformation. Experts emphasize the need for enhanced automated deepfake detectors to counter this evolving threat.

AI in Asia

How China is fast-tracking high-speed rail with AI-powered builders: China railway experts are utilizing AI-powered robots to construct labor-intensive components of the world’s largest high-speed railway network, leading to faster, safer, and more efficient construction. The advancements in robotic construction have broader implications for infrastructure projects within China and worldwide. The integration of AI in high-speed rail construction demonstrates China’s commitment to technological innovation and its potential impact on the global transportation sector.

Toyota, Pony.ai plan to mass produce robotaxis in China: Toyota and Pony.ai are planning to establish a joint venture to mass-produce robotaxis in China. The venture will use Pony.ai’s autonomous driving technology and ride-hailing services. Toyota and GAC-Toyota, a joint venture with Guangzhou Automobile Group, will invest over 1 billion yuan ($140 million). The move aligns with Toyota’s focus on smart cockpits and electric vehicles for the Chinese market. Pony.ai has been active in the self-driving car space, offering robotaxi services in multiple Chinese cities and the United States.

AI in Science and Health

Scientists develop AI system to alert us of next pandemic: Scientists from Scripps Research and Northwestern University have developed an AI system called early warning anomaly detection (EWAD) that utilizes machine learning to predict the emergence of dangerous virus variants in future pandemics. Through analyzing genetic sequences, frequencies, and mortality rates of virus variants, EWAD accurately predicted variants of concern (VOCs) during the COVID-19 pandemic. The system’s ability to anticipate threats before they are officially designated by the World Health Organization could revolutionize pandemic preparedness and aid in understanding viral evolution for better treatments and prevention strategies.

AI could speed up formation of material laws: Binghamton University’s Assistant Professor Pu Zhang has secured a National Science Foundation grant to explore how AI can expedite the discovery of complex material laws. Traditional formula creation for material properties can take decades, but Zhang’s research aims to use AI to analyze and interpret patterns from raw data, leading to faster discoveries. Employing symbolic AI, which interprets and generates equations, Zhang hopes to avoid common AI pitfalls, such as “hallucinations” or incorrect responses. The project may potentially revolutionize science and other scientific fields by harnessing AI’s potential.

AI in Military and Defense

XQ-58 Valkyrie Solves Air Combat 'Challenge Problem' While Under AI Control: The U.S. Air Force successfully tested its XQ-58A Valkyrie drone with advanced autonomous capabilities, using new AI-driven software. The drone completed aerial combat tasks autonomously after being trained millions of times in simulated environments. The test is part of the Next Generation Air Dominance modernization initiative, aiming to develop AI/ML agents for air-to-air and air-to-surface skills. While there will be a human operator in the loop (for now) the Air Force envisions integrating artificially, trained neural networks into real-world operations for future capabilities. China’s PLA is also actively pursuing AI-driven capabilities in tactical air combat.

AI Hardware Market

AMD Confirms Next-Gen Instinct MI400 Series AI Accelerators Already In The Works: AMD has confirmed its plans to develop the next-gen MI400 Instinct series AI accelerators, aiming to compete aggressively in the AI industry. The CEO, Lisa Su, emphasized the need to improve software development for better support of generative AI applications. AMD’s Instinct lineup offers top-of-the-line hardware specs, but they aim to enhance software support to rival NVIDIA. Apart from the MI400 series, AMD also disclosed plans for “cut-down” MI300 variants for Chinese markets. The MI400 accelerators are expected to bring better performance and value, challenging NVIDIA’s dominance in the AI market.

AI chip startup Tenstorrent lands $100M investment from Hyundai and Samsung: AI chip startup Tenstorrent has secured a significant $100 million investment from Hyundai Motor Group and Samsung Catalyst Fund in a convertible note funding round. The funding will be used for product development, design, and machine learning software roadmap. Tenstorrent specializes in AI processors and licenses AI software solutions, and its recent focus on licensing and partnerships may help it compete against industry giants like Nvidia, Google, and Amazon, who are also vying for dominance in the AI chip market. The AI chip sector is predicted to be a significant part of the $450 billion semiconductor market by 2025.

🎧 Did you know AI Breakfast has a podcast read by a human? Join AI Breakfast team member Luke (an actual AI researcher!) as he breaks down the week’s AI news, tools, and research: Listen here

5 new AI-powered tools from around the web

Vmake is an AI-powered editing suite available on iOS and Android, empowering users to create captivating images and videos. Features include video quality enhancement, background removal, object erase, and more alongside AI marketing assistance and downloadable templates for seamless content creation.

ShoppingBotAI is an AI chatboot tool for eCommerce merchants. It allows users to set up a personalized chatbot that can answer customer questions, boost customer satisfaction, and increase sales. It supports multiple languages and offers email customer support.

AI Pet Photos lets you transform your pet into captivating characters with AI-generated, high-resolution portraits in under 2 hours. Create 21 unique designs, from superheroes to scuba drivers, with ease. Safe and print-ready images ensure a hassle-free experience. No design skills needed.

Twon is an AI-tool designed to enhance Twitter (erm… “X”) and Threads posts, providing users with endless post variations in minutes. Utilizing Node-based technology. Twon helps users explore and optimize their content effectively, promising a seamless and organic writing process.

Code Blocks by Sttabot AI revolutionizes app development offering AI-generated code blocks for frontend, backend, and database components, and streamlining native application creation. Developers can effortlessly deploy applications with speed and efficiency, promising lucrative coding experience.

arXiv is a free online library where scientists share their research papers before they are published. Here are the top AI papers for today.

The paper proposes a smartphone camera app that automatically captures both foreground and background blur effects. It detects and segments the subject, tracks motion, and aligns images for sharpness. An underexposed burst is selected to create controlled blur trails, and inter-frame motion is synthesized for smooth blur. The final image combines blurred and sharp elements for high-resolution and HDR results. Implemented on Google Pixel 6 and 7 smartphones, the system democratizes long-exposure photography, enabling casual users to achieve stunning creative effects in just a few seconds with a simple tap of the shutter button.

This research explores multimodal neurons in pre-trained text-only transformers. The authors augment a frozen text transformer with a vision using a self-supervised visual encoder and a linear projection learned on an image-to-text task. They identify “multimodal neurons” that convert visual representations into corresponding text and find that translation between modalities occurs deeper within the transformer. The study demonstrates that multimodal neurons operated on specific visual concepts and have a causal effect on image captioning. Soft-prompt inputs to the language model do not directly encode interpretable semantics. Understanding these neurons can help investigate how transformers generalize across tasks with multiple modalities.

The paper introduced MusicLDM, an innovative text-to-music generation model that adapts Stable Diffusion and AudioLDM architectures to the music domain. To address data limitations and plagiarism concerns, the authors propose two mix-up strategies: beat-synchronous audio mixup (BAM) and beat-synchronous latent mix-up (BLM). BAM linearly combines two music samples in the feature space, while BLM performs mixup in the latent space, generating new samples within the music manifold. The experiments show that BLM outperforms other approaches in generating high-quality and novel text-to-music outputs.

The All-Seeing Project aims to create a comprehensive system for panoptic visual recognition and understanding of the open world. It introduces the All-Seeing 1B (AS-1B) dataset, comprising over 1 billion region annotations with semantic tags, question-answering pairs, and detailed captions. The dataset covers a wide range of 3.5 million common and rare concepts in the real world. Leveraging this dataset, the All-Seeing Model (ASM) is developed, a unified framework for panoptic visual recognition and understanding. The model is trained with open-ended language prompts and locations, demonstrating impressive zero-shot performance on various vision and language tasks. This project lays the foundation for vision-language AI research.

The paper introduces Compartmentalized Diffusion Models (CDM), enabling the training of separate diffusion models on distinct data sources and composing them at inference time. CDMs facilitate data protection, including selective forgetting, continual learning, and serving customized models based on user access rights. Each model contains information only from the data it was exposed to during training, ensuring data security. The method employs a closed-form expression for the reverse diffusion flow of a mixture distribution. Efficient prompt tuning is used for implementation. CDMs achieve performance comparable to a model trained on all data while providing data protection and attribution capabilities, even for large-scale diffusion models.

Thank you for reading today’s edition.

Your feedback is valuable.


Respond to this email and tell us how you think we could add more value to this newsletter.