Trend Micro envisages a new era as cybercriminals use machine learning to create advanced threats

Cybersecurity companies use machine learning technology to enhance threat detection capabilities that help fortify organizations’ defense against malware, exploit kits, phishing emails, and even previously unknown threats, Trend Micro revealed. 

The Capgemini Research Institute conducted a study on the usage of machine learning for security and found that of 850 senior executive respondents based in 10 countries, around 20 percent started using the technology before 2019, and about 60 percent will be using it by year’s end.

The use of machine learning in cybersecurity — not to mention in many other fields across various industries — has proven to be beneficial. This technology, however, is also at risk of being used by threat actors. While widespread machine learning weaponization may still be far off, research concerning this area, particularly the use of deepfake technology to extort and misinform, have recently become a topic of interest for the IT community and the general public.

To get a clearer picture of what is possible and what is already a reality with regard to machine learning-powered cyberthreats, research on machine learning-powered malware is still surprisingly scarce, considering that some experts have long considered it as a type of threat that can possess advanced capabilities. In fact, only one PoC of such a threat has been publicized, which was unveiled at Black Hat USA 2018. 

IBM presented a variant named DeepLocker that can deploy untraceable malicious applications within a benign data payload. The malware variant is supported by deep neural networks (DNN) or deep learning, a form of machine learning. The use of DNN disguises the malware’s conditions, which are pieces of information that security solutions need to detect malicious payload.

DeepLocker is designed to hide until it detects a specific victim. In the demonstration, DeepLocker was seen stealthily waiting for a specific action that will trigger its ransomware payload. The action that triggered the payload was the body movement of a targeted victim when he/she directly looked at a laptop webcam, which is operated by a webcam application embedded with malicious code. The application of machine learning in this attack can be considered limited, but it showed how malware variants can be highly evasive and targeted when infused with machine learning.

Experts are increasingly warning the public about deepfake videos, which are fake or altered clips that contain hyperrealistic images. Produced from generative adversarial networks (GANs) that generate new images from existing datasets of images, deepfake videos can challenge people’s perception of realities, confusing our ability to discern what is true from false.

Deepfake technology is mostly used in videos involving pornography, political propaganda, and satire. As for the impact of these videos, a Medium article published in May claimed that there were about 10,000 deepfake videos online, with House Speaker Nancy Pelosi and Hollywood star Scarlett Johansson being two of their popular subjects/victims.

Regarding the use of this technology in profit-driven cybercrime, it can be surmised that deepfake videos may be used to create a new variation of business email compromise (BEC) or CEO fraud. In this variation of the scheme, a deepfake video can be used as one of its social engineering components to further deceive victims.

But the first reported use of deepfake technology in CEO fraud came in the form of audio. In September, a scam involving a deepfake audio was used to trick a U.K.-based executive into wiring US$243,000 to a fraudulently set-up account. The victim company’s insurance firm stated that the voice heard on the phone call was able to imitate not only the voice of the executive being spoofed, but also the tonality, punctuation, and accent of the latter.

Brute force and social engineering methods are old but popular techniques that cybercriminals use to steal passwords and hack user accounts. New ways to do this could be inadvertently aided by user information shared on social media – some still embed publicly shared information into their account passwords. Additionally, machine learning research on password cracking is an area of concern that users and enterprises should closely pay attention to.

Back in 2017, one of the early proofs of machine learning’s susceptibility to abuse was publicized in the form of PassGAN — a program that can generate high-quality password guesses. Using a GAN of two machine learning systems, experts from the Stevens Institute of Technology, New Jersey, USA, were able to use the program to guess more user account passwords than popular password cracking tools HashCat and John the Ripper.

To compare PassGAN with HashCat and John the Ripper, the developers fed their machine learning system more than 32 million passwords collected from the 2010 RockYou data breach, and let it generate millions of new passwords. Subsequently, it attempted to use these passwords to crack a hashed list of passwords taken from the 2016 LinkedIn data breach.

The results came back with PassGAN generating 12 percent of the passwords in the LinkedIn set, while the other tools generated between 6 percent and 23 percent. But when PassGAN and Hashcat were combined, 27 percent  of the passwords from the LinkedIn set were cracked. If cybercriminals are able to devise a similar or enhanced version of this methodology, it could be a potentially reliable way to hijack user accounts.

Adversarial machine learning is a technique that threat actors can use to cause a machine learning model to malfunction. They can do so by crafting adversarial samples, which are modified input fed to the machine learning system to mess up its ability to predict accurately. In essence, this technique — also called an adversarial attack — turns the machine learning system against itself and the organization running it.

This method has been proven capable of causing machine learning models for security to perform poorly, for example, by making them produce higher false positive rates. They can do this by injecting malware samples that are similar to benign files to poison machine learning training sets.

Machine learning models used for security can also be tricked using infected benign Portable Executable (PE) files or a benign source code compiled with malicious code. This technique can make a malware sample appear benign to models, preventing security solutions from accurately detecting it as malicious since its structure is still mostly comprised of the original benign file.

When it comes to dealing with advanced password cracking tools such as the machine learning-powered PassGAN, users and organizations can move towards two-factor authentication schemes to reduce their reliance on passwords. One approach to this is using a one-time password (OTP) — an automatically generated string of characters that authenticates the user for a single login session or transaction.

Meanwhile, technologies are continuously being developed to defend against deepfakes. To detect deepfake videos, experts from projects initiated by the Pentagon and SRI International are feeding samples of real and deepfake videos to computers. This way, computers can be trained to detect fakes. 

To detect deepfake audio, experts are training computers to recognize visual inconsistencies. And as for the platforms where deepfakes can creep in, Facebook, Google, and Amazon, among other organizations, are joining forces to detect them via the DeepFake Detection Challenge (DFDC) — a project that invites people around the world to build technologies that can help detect deepfakes and other forms of manipulated media.

Adversarial attacks, on the other hand, can be prevented by making machine learning systems more robust. This can be done in two steps: First, by spotting potential security holes early on in its design phase and making every parameter accurate, and second, by retraining models via generating adversarial samples and using them to enhance the efficiency of the machine learning system. 

Reducing the attack surface of the machine learning system can also ward off adversarial attacks. Since cybercriminals modify samples in order to probe a machine learning system, cloud-based solutions, such as products with Trend Micro XGen security, can be used to detect and block malicious probing.

Governments and private organizations, particularly cybersecurity companies, should anticipate a new era where cybercriminals use advanced technologies such as machine learning to power their attacks. As they have done in the past, cybercriminals will continue to develop more advanced and new forms of threats to be one step ahead. In this light, technologies for combating these threats should likewise continue to evolve. 

However, while it would be a good choice to implement a tailor-fit technology to detect such threats, a multilayered security defense (one that combines a variety of technologies) and the consistent application of cybersecurity best practices are still the most effective ways to defend against a wide range of threats.

Samsung is developing AI ScaleNet capability, enabling seamless, high-resolution 8K streaming

As the world enters the era of 8K TVs with sales of 8K displays steadily increasing, TV manufacturers are constantly adding more offerings to the mix, the 8K market is expected to continue to grow. 

There are, however, a few challenges that need to be addressed before viewers around the world will be able to enjoy 8K’s stunning visuals in their entirety. First, more 8K content will need to be produced, and second, network connections need to be made capable of supporting the seamless streaming of 8K movies and shows.

To address these issues, researchers from Samsung Research, an advanced R&D hub within Samsung Electronics’ SET Business, have developed an AI Codec known as AI ScaleNet. First introduced at SDC19, this capability enables delivery of 8K content on networks that typically support only 4K speeds, without the need for additional infrastructure.

To learn more about AI ScaleNet’s development process and potential applications, Samsung Newsroom interviewed Kwangpyo Choi and Youngo Park – researchers from Samsung Research’s Visual Technology team, and the first individuals to suggest and develop this type of innovative technology.

AI ScaleNet utilizes deep learning technology to minimize data loss during compression and enable 8K content to be streamed on networks with lower bandwidth capabilities.

Here’s how it works, in a nutshell: 8K content is compressed to 4K quality using an AI downscaler and transmitted to the user’s TV, which utilizes AI to upscale the content back to 8K quality. When asked to describe the inspiration for the technology, Kwangpyo Choi, a researcher from Samsung Research’s Visual Technology, specified a need for a new video compression and transmission technology in order to accommodate the evolving media landscape, and to address related technical challenges.

“Multimedia content is rapidly moving online, so figuring out how to overcome bandwidth limitations for UHD content transmission has become a major task,” said Choi.

“With AI ScaleNet, users will be able to enjoy 8K-quality content, even at lower bandwidths,” added Youngo Park, another Samsung researcher. “In fact, during periods of high network traffic, when speeds tend to drop, viewers can still expect content to be presented at a relatively higher quality.”

AI ScaleNet addresses bandwidth limitations within modern network infrastructures with what’s known as an adaptive AI Codec. “AI Adaptive Bit Rate Streaming refers to AI technology that adapts to various bandwidths and adjusts the resolution to enable seamless streaming,” said Choi.

The difference the technology makes is especially clear when using over-the-top (OTT) services, where viewers typically notice changes in picture quality. As Choi explained, however, Samsung’s AI Codec and AI ScaleNet technology utilize AI that adapts to changes in bandwidth and optimizes quality, so users enjoy the highest-quality streaming experience possible.

Samsung Research Centers earn top places in leading AI challenges

Samsung Electronics’ Global Research & Development (R&D) centers play a key part in developing artificial intelligence (AI) capabilities for real-world usage. A credit to the work this advanced R&D branch of Samsung undertakes, both Samsung R&D Institute Poland and Samsung Research America AI Center have recently won two prestigious global challenges.

2019 marks the third year in a row that Samsung R&D Institute Poland, in partnership with the U.K.’s University of Edinburgh (UEDIN), has received accolades at the International Workshop on Spoken Language Translation (IWSLT), one of the top two global workshops on automatic language translation, along with the Workshop on Machine Translation (WMT). 

This year, Samsung R&D Institute Poland won first place in two categories, the first being text-to-text translation from English to Czech and the second – an end-to-end system translating English speech into German text.

For the text-to-text translation category, researchers worked to develop a model to translate the transcript of a spoken English-language TED Talk into Czech. Developing their winning model required the Samsung team to develop large, filtered corpora from which to work and generate as much synthetic data as possible. 

The work done by the Samsung R&D Institute Poland team, together with additional modeling help from UEDIN, was selected as the best in the challenge by human evaluators. This means that the translations produced by Samsung R&D Institute Poland’s system scored the highest both in fluency and adequacy.

Samsung R&D Institute Poland’s participation in their second winning category this year, the end-to-end translation system from English to German, was a first for the team. 

The task was to produce a German-language transcription of an English-language TED Talk audio recording. This task required the development of a single model that could take an audio file input and subsequently produce a translated transcription. It was made more difficult by the deficiency of the provided audio sources, compared to typical speech recognition task. 

Samsung R&D Institute Poland proposed several innovative methods for end-to-end speech translation that mitigated this source paucity, obtaining a state-of-the-art result with their final system that won them first place in the challenge.

This October, researchers from Samsung Research America’s AI Center received first place in the International Conference on Computer Vision (ICCV)’s challenge: Linguistic Meets Image and Video Retrieval (Fashion IQ). ICCV is a premier international computer vision conference that took place in Seoul, Korea, this year.

The challenge Samsung Research America AI Center took part in, sponsored by IBM research, aims to develop conversational shopping assistants that are more natural and real-world applicable. The task given to Samsung Research America AI Center‘s team, the ‘Superraptors’, belonged to the domain of image retrieval. 

In the task, an input query was specified in the form of a candidate image as well as in two natural language expressions that describe the visual differences of the search target. The goal of this challenge was to gather opinions and experience from researchers on the emerging space of visual content retrieval with a natural language interface.

Samsung Research America AI Center’s submission to the challenge, “Multimodal Ensemble of Diverse Models for Image Retrieval Using Natural Language Feedback”, blended the given data in different modalities with multiple deep learning models. 

The team’s win marks the first time a Samsung Research team has won a multimodal (language and vision) challenge; previously, Samsung AI Center Moscow, Samsung R&D Institute Poland and Samsung R&D Institute China-Beijing have received awards in single modality challenges.

Supermicro aligns with Intel to offer large scale distributed training AI systems to meet needs of deep learning training models

Super Micro Computer is collaborating with Intel on AI solutions that are validated on the Intel Nervana Neural Network Processor for Training (NNP-T). Intel's NNP-T is a purpose-built AI training ASIC supporting the growing compute needs of deep learning training models.

The Intel Nervana NNP-T solves memory constraints and is designed to scale out through systems with racks easier than today's solutions. As part of the validation process, Supermicro integrated 8 NNP-T processors, dual 2nd Generation Intel Xeon scalable processors, up to 6TB DDR4 memory per node supporting both PCIe card and OAM form factors. Supermicro NNP-T systems are expected to be available mid-year 2020.

With high compute utilization and high-efficiency memory architecture for complex deep learning models, the Supermicro NNP-T AI System is built to validate two key real-world considerations: accelerating the time to train ever-complex AI models and doing it within a given power budget. The system enables faster AI model training with images and speech, more efficient gas & oil exploration, more accurate medical image analytics, and faster autonomous driving model generation.

"Supermicro is excited to cooperate with Intel for the Nervana NNP-T platform," said Charles Liang, president, and CEO of Supermicro. "Striking a balance among computing, communication, and memory, the validated NNP-T ASICs on Supermicro systems can train large AI models with near-linear scaling efficiency via intra- and inter-chassis links."

"Supermicro has validated our Deep Learning (DL) solution and is helping us prove the Nervana NNP-T system architecture, including card and server design, interconnect, and rack," said Naveen Rao, corporate vice president and general manager, Artificial Intelligence Products Group, Intel.

Microsoft Research shows that logarithmic mapping allows for low discount factors by creating action gaps similar in size

While reinforcement learning (RL) has seen significant successes over the past few years, modern deep RL methods are often criticized for how sensitive they are with respect to their hyper-parameters. One such hyper-parameter is the discount factor, which controls how future rewards are weighted compared to immediate rewards.

The objective that one wants to optimize in RL is often best described as an undiscounted sum of rewards (for example, maximizing the total score in a game). 

In practice, however, a discount factor is introduced to avoid some of the optimization challenges that can occur when directly optimizing on an undiscounted objective. And while in theory a discount factor can take on any value between 0 and 1, in reality good performance is only obtained for a small subset of values close to 1.

Microsoft Research will present a paper at the thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019), called “Using a Logarithmic Mapping to Enable Lower Discount Factors in Reinforcement Learning.” 

The company will present a novel hypothesis as to why the effective discount-factor range is so small. Also, based on its hypothesis, Microsoft outlined a technique to avoid this discount-factor sensitivity and introduce a method, which is called logarithmic Q-learning, based on this technique. 

Logarithmic Q-learning is the first method able to achieve good performance for low discount factors on sparse-reward tasks with function approximation. In addition, the method not only reduces discount-factor sensitivity, but can also improve performance altogether.

While a logarithmic mapping function makes action gaps more homogenous, there are a number of challenges to overcome to build a robust algorithm, based on this outcome, that can be applied to stochastic domains with both positive and negative rewards. 

Microsoft managed to overcome these challenges and develop an algorithm, logarithmic Q-learning, that can be applied to general tasks and has convergence guarantees under standard conditions. These results do not only show great performance for low discount factors. Early performance of high discount factors is also better than it was without the logarithmic mapping.