Cathal Smyth

Prompt engineering has emerged as a critical skill for effectively interacting with large language models. This comprehensive study examines the diverse landscape of prompting techniques, security vulnerabilities, and key considerations for prompt design. The authors have done a great job creating a taxonomical structure of prompt techniques, making it easier for AI practitioners to systematically implement and understand the relationships among them. A highlight for me was the detailed case study on using prompt engineering to detect signals of suicidal crisis in text, showcasing a step-by-step process to improve prompts and enhance results. The study also emphasizes how sensitive LLM responses are to different prompts and the black-box nature of these models, underscoring the need for automated prompting techniques to minimize result variability. Paper - https://lnkd.in/gjHtXmDD #LLM #AI #PromptEngineering

29

Recently, the Cyber Safety Review Board determined that NIST 800-53 requires updates to address better the risks associated with cloud networking. With Canada's ITSG-33 regarded as the equivalent security framework, does this mean the Government of Canada's baseline information technology security framework insufficiently protects Government networks? #cloud #cloudsecurity #cybersecurity #cyberdefence #ITSG33 #NIST80053 #informationsecurity #canada #innovation #nationalsecurity #enterprise

4

1 Comment

In our latest paper with Jean-Baptiste Waring, B.Eng., we generated a dataset on `ibm_quebec` the 127-qubit IBM Quantum System One located in Canada. The dataset contains more than 4000 experiments of long-range entanglement path qubits from 0 to 100+ qubits. The dataset and the code can be found here: https://lnkd.in/eussBW9G Paper: https://lnkd.in/ekb9Vqja A special congratulations to Jean-Baptiste Waring, B.Eng. for this repo and the quality of the experiments. #quantumcomputing

47

1 Comment

Excellent remarks to Senate Committee on National Security, Defence and Veterans Affairs in Canada by Social Media Lab. The conclusion is extremely, and enduringly relevant to New Zealand's information, and media environment as one which closely mirrors Canada's context, but without any comparable official interest, investment or investigation whatsoever into truth decay, (foreign) malign influence operations or information integrity. "In conclusion, we must not underestimate the potential of the Kremlin’s information operations to impact public perception over time. Deplatforming individual sources may not be effective as it could also undermine trust in the government and legitimize censorship. A more nuanced approach should consider the various forms of interference and develop strategies to address them directly. This could include requiring large social media platforms to expand their Trust and Safety teams in Canada, share data with researchers to promote transparency and allow for independent audits. Schools must update digital literacy programs to address emerging challenges such as Generative AI. For the general public, the government should develop PSAs and pre-bunking campaigns to educate Canadians about foreign inference. Aspects of such educational campaigns must focus on and be informed by diaspora communities in Canada, as they are more likely to be targeted by foreign states." These are vital recommendations to consider in concert as bulwarks against what is an unceasing, and increasingly sophisticated assemblage of beguilingly compelling narratives that undermine democracy, eviscerate electoral integrity, and seamlessly cross borders. It's what I study every day.

4

The authors shown below put out a well-read academic paper on Canadian corruption issues just a few months back. That this issue continues to be highlighted by CSIS, CISC, RCMP, foreign intelligence services, and now clearly stated in academic papers... and yet is completely ignored by government... is disastrous for our nation. The well-documented complaints of thousands of citizens is denied, delayed & ignored in practiced patterns to frustrate fairness & just responses, leading to ever increasing corruption & infiltration by organized crime. Even the firing of our Attorney General, Jody Wilson-Raybould, was a cover-up of massive corruption issues in Canada... and was accomplished without any repercussions for the perpetrators. My own experience of having millions of dollars of R&D crushed by career bureaucrats has been refused any fair review even though 4 Federal Ministers had requested the tech, US Homeland had vetted the tech and approved it for BC/Alaska border testing, and CBSA & CRA documentation clearly states that there was no rationale to attack this company that was providing leading-edge border security enhancement... unless it was to defend illicit trade & money laundering. Below are excerpts from the academic paper. The conclusions they make are ignored at our own peril: The impact of neoliberal policies has been profound. The destruction of regulation, regulatory agencies and regulatory agents through downsizing, defunding and deregulation removed many of those responsible for passing and enforcing laws punishing corporate wrongdoing. While deregulation in the United States started with the election of Ronald Reagan in 1980, Donald Trump carried it to an extreme. While successive Canadian governments have never directly copied their U.S. counterparts, we, too, have witnessed the negative outcomes associated with globalization. Given all this, serious action against corruption, in Canada and abroad, must include moving beyond the narrow and reactive confines of bribery law and policy, as espoused by the OECD and woefully mishandled by the Canadian government, to confront the harms associated with globalization and bring multinational corporations under democratic control.

4

6 Comments

👉🏼 ChatENT: Augmented Large Language Model for Expert Knowledge Retrieval in Otolaryngology-Head and Neck Surgery 🤓 Cai Long 👇🏻 https://lnkd.in/e7_3F-hi 🔍 Focus on data insights: - ChatENT outperformed ChatGPT4.0 in both the Canadian Royal College OHNS sample examination questions challenge and the US board practice questions challenge. - A significant error reduction of 58.4% and 26.0% was observed, respectively. - ChatENT demonstrated enhanced performance in analyzing and interpreting OHNS information, generating fewer hallucinations and showing greater consistency. 💡 Main outcomes and implications: - ChatENT is the first specialty-specific knowledge retrieval artificial intelligence in the medical field utilizing the latest LLM. - Promising applications in medical education, patient education, and clinical decision support are highlighted. - Overcoming limitations of existing LLMs, ChatENT signals a future of more precise, safe, and user-friendly AI applications in OHNS and other medical fields. 📚 Field significance: - Artificial Intelligence - Medical Education - Clinical Decision Support 🗄️: [#Otolaryngology #ArtificialIntelligence #MedicalEducation #ClinicalDecisionSupport]

1

To what extent do Large Language Models (LLMs) remember their training data? Can we detect whether a model has been trained on particular data samples? What does that imply in terms of copyrighted content watermarking, and data privacy? Our work "Copyright Traps for Large Language Models" that we are presenting this week at ICML in Vienna attempts to tackle these questions! ➡️ A little known fact about 🥐 CroissantLLM is that we inserted a few special text sequences amongst the 3 trillion token training set in order to study model memorization in a rare, randomized controlled experimental setup. These "trap" sequences represent less than 0.04% of the data, and are designed to be unlikely to be filtered out by dataset preprocessing methods. Measuring the extent to which CroissantLLM "remembers" these sequences enables us to better understand the settings under which memorization occurs. Typically, we find that "trap" sequences that have been repeated in the training set a higher number of times, that are lengthier, or that are more "surprising" tend to be easier to detect. 💡 What does this mean? For CroissantLLM, in which internet data is never repeated more than 4 times, it reassures on the fact no individual sequence has a real chance of being remembered by the model. Thus, no risk of the model leaking out social security numbers or other sensitive information that slipped into the training set! 🪤 Focusing on the problem the other way around, it also means data owners could design and insert "copyright traps" within their data, to be able to accurately detect whether it has been used to train language models! For more information, don't hesitate to go check out: 📜 Our paper: https://lnkd.in/ePJArTJB 📰 The CroissantLLM report: https://lnkd.in/eM2Yx6ZE 🪧 For those attending #ICML, the poster session on July 23rd! This work is a collaboration with Matthieu Meeus, Igor Shilov and Yves-Alexandre de Montjoye from the Computational Privacy Group of Imperial College London ! I was super glad to also be able to explore these privacy-related aspects of LLMs as part of my PhD with CentraleSupélec and ILLUIN Technology in which I generally focus on more industrial applications.

85

3 Comments

The April Rewire is now out to help you navigate some big cyber defence topics from April. I talk about the APT31's targeting of Canada's Members of Parliament and more. https://lnkd.in/gKhamFGH #Canada #cyberdefence #cybersecurity #apt31 #houseofcommons #canadianarmedforces #innovation #technology Canadian Cyber in Context

5

Not surprising ... promising theoretical ideas don't always work out in the end. That's how research works. Potential attacks get posted from time to time, and it's best to not panic. The world of quantum safe algorithms is safe again, for now :) https://lnkd.in/ev-A4X3r

11

I'm excited to finally publish the last article in my dissertation! Key Contributions: (1) Performing the first known study of resource-efficient machine learning–assisted software vulnerability detection; (2) Presenting Vul-Mixer, a resource-efficient architecture inspired by how the human brain processes code; and, (3) Demonstrating Vul-Mixer maintains 98.3% of SOTA with only 0.2% of the parameters and 173MB of memory. There's a lot of theory behind *why* we were able to get these results and I believe it points to broader problems with the datasets and the structure of the task itself. But I'm happy to have this work out there! #cybersecurity #vulnerabilitydetection #mlavd #mdpi #mlsec #aisec https://lnkd.in/eRAws8fv

32

3 Comments

This is a huge problem. Every risk assessment available indicates that this is a top risk. Yet the Government of Canada falls short according to this audit by the Office of the Auditor General of Canada / Bureau du vérificateur général du Canada https://lnkd.in/gsTCx6S5 #ITsecurity #it #cybercrime #itaudit #audit #performanceaudit

2

Great news in the world of #trustworthyAI - the ARIA program by #NIST - a much needed evaluation tool for assessing the risks and impacts of AI. ARIA expands on the AI Risk Management Framework, released by NIST in January 2023 The program uses 3 measurements to enhance our understanding of AI models, their capabilities, potential risks, harmful impacts, and beneficial applications: 1. Model Testing - helps identify the AI model's capabilities 2. Red Teaming - helps identify worst case scenarios, thereby highlighting risks that must be addressed before real-world deployment. 3. Field Testing - helps identify how users interact with the model and its outputs in the real world. Thank you for sharing Reva Schwartz! #ethicalAI, #ai

10

The Treasury Board of Canada Secretariat | Secrétariat du Conseil du Trésor du Canada has just published the Government of Canada's Enterprise #Cyber #Security Strategy. Great to see a focus on #resilience as well as a push to improve how departments and agencies respond to cyber events - as I have said before, it is not a matter of if but when and how bad when it comes to a cyber incident. I also like the focus on identity and access management that is in this strategy - a critical element in our modern, distributed technology environments. And it was good to see the issue of third party risks highlighted as well as the #talent challenges in cyber security. But what I really liked was in Appendix A, where it looks like the strategy will have many quantifiable KPIs and I look forward to seeing how these indicators develop over time. This seems like a good start - thanks Po Tea-Duncan for sharing this. #GCDigital #Infosec https://lnkd.in/gRzWkntq

14

1 Comment

Many have questioned Kyber being selected as a standard for PQC (Post Quantum Cryptography) Key Exchange, especially in the way that it struggles to scale over the standard parameter sets. Personally, I like FrodoKEM, as it's a purer version of LWE (Learning with Errors): https://lnkd.in/ej7wQuvx

1

1 Comment

Ontario's Bill 194 "An Act to enact the Enhancing Digital Security and Trust Act, 2024 and to make amendments to the Freedom of Information and Protection of Privacy Act respecting privacy protection measures" passed the 1st Reading! The Bill includes - the creation of a new statute, the Enhancing Digital Security and Trust Act (EDSTA), and -amendments to the Freedom of Information and Protection of Privacy Act (FIPPA). - provides new tools to prevent and respond to cyber security threats and safeguard critical public services, such as health care and education. - strengthens safeguards for children’s personal information - lays the foundation for the ethical use of artificial intelligence in the public sector - increases the authority of the Information and Privacy Commissioner of Ontario (IPC) to investigate and respond to privacy breaches. #Bill194 #Cybersecurity #Privacy #Dataprotection #AIethics #Digitalsecurity #Digitalrights

1

Check out Government of British Columbia's "Security Day: Guardians of the Cloud" live today! Sign up closes shortly: https://lnkd.in/gbJjs_b4 on Eventbrite You can catch Louise Turner and I from Quantum Algorithms Institute at 10amPST/1pmEST introducing #Canadian #cybersecurity to #quantum technologies and how they will influence #informationsecurity and #encryption in the coming years.

Software debloating is a promising technique not only for managing attack surfaces but also bringing the potential for managing software dependencies and securing the software supply chain. In this work, we studied the literature in this domain, developed a multi-level taxonomy, and discussed future research directions. The following are key aspects: - We collected 48 academic papers published over the years 2000 - 2024 at software engineering and security venues - Generating SBOM, CI/CD integration, leveraging ML/LLM to facilitate software debloating, and exploring the impact of debloating on sustainability are all future research directions. This work is in collaboration with Yazan Boshmaf and Ashish Gehani 📜 https://lnkd.in/gDGD4_Pg

6

1 Comment

The Government of Canada released an Enterprise Cyber Security Strategy last month. Unfortunately, most media outlets overlooked or did not understand what enterprise cyber security means or is about. So, I decided to write an introduction to the first Canadian Enterprise Cyber Security Strategy. https://lnkd.in/gQ4nP7dc #Canada #defence #defenceindustrialbase #enterprise #cybersecurity #cyberdefence Canadian Cyber in Context #innovation #cyber #treasuryboard

24

8 Comments

Researchers with Google say they uncovered and shut down a number of government disinformation campaigns. The team at the Google Threat Analysis Group (TAG) reported shutting down multiple accounts across its services that were targeting both their domestic and foreign governments seeking to influence public opinion. The report, which covers the Q2 period, detailed account takedowns it made around disinformation and fraud campaigns. It also acts as something of a "canary" report which allows the TAG team to indirectly hint at larger efforts by oppressive governments to manipulate their populations via social media. Among the most prolific violators was Indonesia, which logged two instances in which a total of 39 accounts were shut down for posting misinformation on behalf of the country’s ruling party. According to monitoring site Freedom House, Indonesia has had a fairly spotty record when it comes to internet freedom. The country scored a 46 on its Freedom of the Net rankings. For comparison, the U.S. scored 76. The single largest takedown of accounts went to China, where YouTube administrators managed to wipe out 1,320 accounts, and Blogger took down 1,177 blogs related to what it called a public influence campaign regarding diplomatic relations between China and the U.S. Also making the TAG report was Pakistan, where a total of 59 YouTube accounts had to be pulled down for posting content that targeted politicians in that country with disinformation content. Russia, not to be outdone, also managed to win itself a massive takedown. Google said it removed a whopping 378 YouTube accounts in the quarter linked to disinformation. “The campaign was linked to a Russian consulting firm and was sharing content in Russian that was supportive of Russia and critical of Ukraine and the West,” said Google Tag, also noting a second Russian campaign in which a further 11 YouTube accounts were taken offline for posting misinformation related to Ukraine. Not every one of the disinformation campaigns was performed on behalf of the government, the TAG report noted that Google also busted up a disinformation campaign based out of India and the Philippines. In that case, bogus news was being pushed out in English and Norwegian in hopes of making a quick buck off of advertising. https://lnkd.in/gDprm6sT

11

2 Comments