The abilities of smart technologies are huge and that is ways it is also both important and difficult to control the risks associated with it.
Analytics India Magazine, November 10, 2020
Coming from Greek steganographia, steganography means concealed or covered. It is the practice of concealing information inside other files, messages, etc. and dates back to the 15th century when secret messages were literally hidden in the physical world. Nowadays the goal of stenography is to carry and hide some digital message. The carrier itself may stay visible, and the message can be encrypted.
Smart technologies are widely used for these purposes nowadays with their ability to hide an image within another image. Still, it is not that easy when it comes to speech data.
According to the researchers, who studied the application of deep neural networks as steganographic functions for speech data, hiding speech instead of text enables preservation of such content, as speaker’s identity. The architecture consisted of (i) Encoder Network (ii) Carrier Decoder Network and (iii) Message Decoder Network with each component implemented as a gated convolutional neural network. The researchers evaluated the approach on TIMIT and YOHO datasets using the standard train/val/test splits to assess the model under various recording conditions.During the study, both human and automatic evaluations were conducted: 400 people’s answers were recorded. In the majority of cases, listeners were able to determine if the speaker in the forth record matched the speaker in the first three. The research showed the effectiveness of the proposed method and demonstrated that the approach could be used for concealing several messages in one carrier.
Analytics Insight, November 15, 2020
If we take Gartner’s Top 10 Strategic Technology Trends for 2020 (published October 2019), by 2022, 30% of all cyberattacks on smart systems will be either training-data poisoning, AI model theft, or adversarial samples.
By analogy with any other popular technology, nowadays cybercriminals exploit loopholes AI systems more and more launching threats or malwares, and the number of cases where malefactors attack smart systems is going to increase in the future. An adversarial attack aims to affect the behaviour of an ML model and is based on the feeding of inputs, be it image, text, or voice, to the ML model. As a result, the fooled model should make some kind of a mistake.
Now, backdoor attacks are very popular when aiming to manipulate the behavior of an AI algorithm implanting adversarial vulnerabilities in the model during its training phase. During the attack, the training samples are modified in a way that the presence of a trigger in the input data leads to mistakes in classifications to some target class. Such attacks can cause a lot of trouble for developers if they outsource work on neural networks or build other products on neural networks available online.
Fedscoop, November 5, 2020
The artificial intelligence shop by the Pentagon is busy developing new ways to protect data and models from the new form of cyberattacks, in which an attacker attempts to fool algorithms in order to gain critical information. The Joint Artificial Intelligence Center is working out the methods to stand against adversarial AI with the help of data-sharing and model-sharing.
The JAIC is currently launching 32 AI products touching various areas from predictive maintenance operations and cybersecurity to warfighter health; now, the data training of the systems is most critical and demands extra security.
“The trick is to figure out what tech/products are ready to deploy and in what ‘domain,’” CTO Nand Mulchandani commented. “We think that AI explainability, AI security, AI ethics, and AI testing are all tightly connected and have very tight collaboration between the various groups that are tackling these areas.”
What is Trusted AI? Asking Google to build a route feels like doing the trust fall exercise. Behind me there is a vague understanding of AI and promises that it ...
