Human-centered evaluation of explainable AI applications: a systematic review

Interview: Safran Defense Boss On The Changing Battlefield And AI

PARIS — Safran Electronics & Defense CEO Franck Saudo spoke to Defense News at the Paris Air Show last week about the changing battlefield and use of AI, as well as areas of future growth. Safran is Europe's biggest supplier of military optronics and inertial navigation systems.

Saudo noted two mega trends in defense, one being rising defense budgets with a new emphasis on European sovereignty, and the second the transformation of the battlefield, including greater battlefield transparency, widespread electronic warfare and new objects such as drones.

On the changing battlefield

Syria "really was the conflict where we saw GNSS-denied environment and contested electromagnetic spectrum, for communication or GPS access," while Nagorno-Karabach "was a swarm of drones. And I would say Ukraine is the fusion of both conflicts, adding the return of the war of attrition, which is really particular to Ukraine."

One element of the transformation "definitely is the transparency of the battlefield, between the sensors on the ground, in the air, in space, doubled by open-source data from social networks."

Franck Saudo (L), then the CEO of Safran Helicopter Engines, is pictured in Bangalore, India, in July 2022. Saudo was appointed CEO of the company's Electronics and Defence unit in April 2023. (Manjunath Kiran/AFP via Getty Images)

"This is really what I would say a transformation of the defense demand. On the battlefield, out of the transformation of the battlefield, we have new objects appearing. Drones are one of them, loitering munitions. They require a different trade-off between technology and needs, and cost efficiency."

"Does it mean big programs are off the table? Of course not. Look at the U.S. NGAD Next-Generation Air Dominance, look at the CCA Collaborative Combat Aircraft. Look at the Golden Dome. We are not saying at all that big programs of record are off the table, but what we are saying, alongside we have an evolution of defense demand."

The transformed battlefield requires reinforced protection of existing assets such as tanks, combat vehicles or frigates, for example by adding counter-UAS capability.

Saudo cited frigates in the Red Sea that were initially firing missiles to counter Houthi drones: "Efficient, but not cost efficient." Safran equipped French frigates with a Paseo XLR electro-optical sensor to see drone threats at 40 kilometers away, allowing early identification and giving frigates the cheaper option of using their 76 mm cannon rather than a missile.

"Reinforcing the protection of existing elements, and alongside, capacity, building mass and what I would call new defense, meaning objects with a different trade-off in terms of technology and needs and cost efficiency. This is a transformation of the defense demand."

On supply-side transformation

"We also see a transformation of the supply side with new players, be it startups or commercial players, but also existing players, where to play in the new defense arena it's all about agility, innovation, shortened cycles, being able to do incremental innovation."

Safran took four weeks to implement PASEO XRL on the French frigates, developed its Skyjacker counter-UAS system based on GNSS spoofing in six months, and in Ukraine integrated the AASM Hammer precision-guided munition on Soviet-era MiG-29 and Su-25 aircraft in four months. "That's indeed agility, innovation, which is part of what is required today."

"When I talk about new defense, agility, innovation, part of it is also market players taking risks." Saudo said Safran has self-financing products including drones, loitering munitions and infrared binoculars, even as he repeated that programs of record will continue.

"In the current environment, you need financially to be able to invest. The defense market is moving fast, and it's in transformation. It's about tech and industry: tech innovation, agility, industry capacity to scale up production.

On defense AI and data

The company's artificial-intelligence unit employs 250 people, "the biggest defense and security pool of AI professionals in Europe for production AI. Many people are talking about developing AI, research AI. This is not the business of Safran.AI. It's in production, at the heart of real, combat-proven missions."

Safran in September acquired Preligens, now Safran.AI, which uses AI to analyze images, video and acoustic signals to identify objects of military interest. Saudo said several armed forces around the world use Safran's AI products for surveillance and detection.

"What we stand for is an open architecture, meaning capacity for the AI solutions of Safran to be agnostic to sensors and harbor data coming from different sensors. Safran is an equipment manufacturer, so our DNA is at equipment level, and being open to work with the different players."

"We are really where the mission requires. Part of it is edge AI, so we already have edge-AI solutions for full motion video." Saudo said the company's Euroflir 410 electro-optical system or a PASEO sight can typically use edge AI for data treatment.

"It's also an embarked data hub, a concentrator of data close to the battlefield, which is a second piece. It can be on a frigate, concentrating the different data sources, or on the battlefield, concentrating different sensors."

"Safran is just an enabler for the armed forces, and then it is for them to decide at which echelon the information sort of circles back, knowing that the beauty of data is it can go in many directions. So you can have sort of short loops, but also longer loops with different level of analysis."

"What we know out of Ukraine is the importance of decentralizing the decisions. If the armed forces decentralize the action, it requires different level of data concentration and making sense of the data. We do not believe in one size fits all, where all the data goes from the battlefield to a satellite to another satellite, and then down to a big sort of center."

"It's for the tacticians to decide the way they operate. All we do is we provide enablers, and because our AI architectures are open, because we have this capacity to have both Edge AI and embarked data hub, we then provide optionality for the armed forces, which enables decentralization where they wish."

On growth areas

Saudo sees growth opportunities in Intelligence, Surveillance, Target Acquisition and Reconnaissance, with AI-enhanced ISTAR core for Safran "because it's at the junction of electro-optical and AI."

A second area of growth is resilient or assured position, navigation and timing. "How do you make sure that thanks to the combination of inertial navigation, GNSS, etc, you are immune to spoofing, to jamming, be it for your munitions, be it for your vehicles, be it for your high-value assets?"

Further growth areas are precision-guided munitions, including the Hammer air-launched weapon and missile seekers, as well as drones and loitering munitions, where Safran produces the Lanner 50-kilogram drone and Warbler loitering munition.

"I would add AI solutions as a stand-alone, because we both embed our AI solutions, edge AI and embarked data hub, but we also provide the Safran.AI open architectures for use to integrators to include as a building block in their suites."

Polymorphic Defense: Cybersecurity For An AI-Driven Threat Landscape

Carlo Tortora Brayda, executive chairman of the Tortora Brayda Institute and co-founder of the National AI & Cybersecurity ISAO.

getty

Cybersecurity has been designed with a static mindset. You built your castle, dug out your moat and braced for impact. Attack workflows were predictable thanks to tools like MITRE ATT&CK frameworks.

Now, things are different. As the age of artificial general intelligence (AGI) rapidly approaches, we are already overwhelmed by the power of AI cyberattacks due to their speed, creativity and sophistication. We need unprecedented agility.

Adversaries are now asymmetric and polymorphic in their attack capabilities. Asymmetric implies that adversaries are not evenly matched, and suggests agility and unconventional tactics. Polymorphic means changeable, having shapeshifting qualities.

This is why polymorphic autonomous cyber defense might be the best defense—not building more walls but changing the shape of the battlefield.

The Need For Polymorphic Defense

Polymorphic defense is a radically new evolution in cybersecurity, but it has deep roots in military doctrine. The basic idea is to create an environment that is in continuous transformation.

Cyberspace does not have boundaries, and our security philosophy has to adapt to adversarial asymmetry and AI-driven agility. It has to become more potent through deception, shifting configurations, obfuscation and firing off all kinds of decoys, creating a hall of mirrors for the enemy.

We need to move away from a brick-and-mortar view of cybersecurity and take inspiration from the past when, as technology evolved, the Great Wall of China or Hadrian's Wall became small hindrances rather than insurmountable obstacles.

Inspired by maneuver warfare, modern cybersecurity must be able to act faster than the attackers; reconfigure systems, data locations, tags and network ports; and simulate attack surfaces quicker than ever before. A continually changing defense invalidates the concept of the Cyber Kill Chain.

Consider the centuries-old Russian military tactic, Maskirovka, which highlighted the importance of camouflage and deception. Planting red herrings to mislead the opponent, making the battlefield impossible to follow.

Cyber-polymorphism achieves this digitally. Future systems will generate endless mirages to trick the attackers down endless rabbit holes, exhausting their desire to succeed. Beware, though, that if the prize is big, it will come down to who has the most computing power and the best algorithms. It will boil down to investment.

The only way to realize this level of defense agility will be through orchestrated agentic AI. Most importantly, we must transcend the idea of humans sanctioning each move. There won't be time. In this vision of future cybersecurity, if you put a human in the decision making defense loop, you have checkmated yourself.

Polymorphic cyber defense takes what the U.S. Department of Defense terms "moving target defense" a giant leap further. Instead of reducing the exposure of attack surfaces by frequently changing system configuration, polymorphic cyber defense applies intelligent, AI-driven continuous transformation across your entire digital estate.

How Polymorphic Defense Can Be Adopted

Let's peek over the fence to see what this technology might look like.

Defenders apply tactics once reserved for malware authors in code and binary polymorphism. Address space layout randomization (ASLR) is an established memory protection technique; this can become very effective when applied dynamically across processes, randomizing memory addresses in real time.

Code obfuscation has long been used to protect against reverse engineering, but we can envisage a future where the technique can be applied across the network map. Reconfiguring your network dynamically through software-defined networking obfuscates the topology of the terrain. Honeypots designed to act like flares on warplanes can be spawned off dynamically and unpredictably.

In data terms, rotating encryption keys, using one-time pad and quantum key distribution and shuffling how data is microsegmented and located will be part of confusing the attacker.

Altering API endpoints and changing system responses can wreck automated attack scripts and simulate unstable systems, tricking attackers into thinking the target is already compromised or corrupted.

Orchestrated AI can analyze static and dynamic threat intelligence, detecting attack patterns and shapeshifting systems pre-emptively. Adaptively changing network topology is practical against distributed denial of service attacks.

Polymorphic defense in cloud environments involving containers can support workload shuffling and dynamic access control to maintain tenant isolation and reduce lateral movement. Lightweight forms of polymorphism can also be applied to IoT cybersecurity, rotating communication protocols and encryption keys, enhancing resilience and being mindful of resources simultaneously.

We are inching into this new reality with pioneering developments such as Morphisec, applying real-time binary polymorphism to endpoints, and earlier obfuscation network-level protocols like Scramblesuit, which help to showcase this novel approach.

This form of defense outpaces the attacker by being proactive and polymorphic. It is the only thing that can give the defender the upper hand. For the adversary, the attack becomes too costly and damages the profitability of the criminal enterprise.

The victors of future cyber conflicts will be the ones with the most adaptive and unpredictable architecture.

Realizing Polymorphic Cyber Defense

However, realizing true polymorphic cyber defense systems will not be easy.

Enterprise and industrial systems can be so complex that the demands on expertise are high, and delivering that level of orchestration will remain immensely challenging.

We can envisage latency, false positives and unchaining needless internal sequences of cyber adaptation. There will be governance issues, and frequent changes to data handling and architectures can make auditability for compliance purposes more difficult. Polymorphism will need to provide GDPR transparency.

These are all obstacles that innovators and inventors will need to overcome. It will take time for such systems to earn the trust of cybersecurity leaders and their boards.

However, with AI progressing at such a rate, realizing true polymorphic defense systems is inevitable, as is their market adoption. The double-edged sword of quantum computing can already provide for true randomness generation, removing the opponent's ability to have a predictable attack algorithm.

CISOs and boards need to support and embrace this technology that symbolizes the best innovation in the field of cyber defense. To be in a position to defeat attackers through the adaptive elegance of an Aikido move.

The future of cybersecurity is polymorphic, agentic and categorically AI-driven. Let's give the bad guys a hard time.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

AI Companies See Openings To Improve Federal, Defense Operations

Large artificial intelligence providers are angling for their foundation models to form the bedrock of government AI use cases, with leaders from both OpenAI and Anthropic pitching the technologies as an expeditious solution for the public sector, namely defense operations.

Following the June 16 announcement of a new OpenAI for Government offering, Katrina Mulligan — the national security head at OpenAI for Government — said the new initiative will focus on tailoring AI tools to innovate in specific public sector operations

"OpenAI is laser-focused on … being the leading foundation model developer," Mulligan said on Thursday at the Defense One Tech Summit. "I really think that there's going to have to be a lot of creativity in how we do this. It's not going to be one way. It's not going to be one uniform application. It's going to be very dependent on the use case, and very dependent on the trade offs that you're willing to make and not willing to make." 

Mulligan added that the gap OpenAI for Government is looking to bridge is between available technology and what is occurring inside government operations, particularly in the defense sector. Beginning its public sector operations with a $200 million pilot program with the U.S. Department of Defense's Chief Digital and Artificial Intelligence Office, she said the best way to understand how to improve operations with AI is through practice and usage.

"We are going to have really great insights about how this technology can transform defense missions," Mulligan said. "But the most brilliant ideas that I think are going to work are gonna emerge from the men and women who serve. And the only way that those ideas are gonna surface is by getting the technology into the hands of those men and women."

Mulligan added that OpenAI's foray into public sector AI implementation is part of the larger modernization agenda that federal agencies have undertaken, particularly in regards to the reduction of mundane and bureaucratic tasks. 

"If you think about what it is like for the men and women who serve and where the toil is in their you know daily workflow, or the toil of a cybersecurity analyst, for example, you can take these jobs that are pretty grueling –– especially at the earlier stages –– and take out the pieces that people really don't enjoy and free them up to do things that only humans can do," she said. 

Mulligan's comments come as other private sector players are also angling to make their AI products available for federal customers. On Wednesday, Anthropic Co-Founder Jack Clark testified before the House China Select Committee that one pillar in America leading AI innovation worldwide is the adoption of AI in national security operations. 

"By securely integrating AI capabilities directly into national security workflows, the U.S. Government can achieve a dual strategic advantage: developing expertise in leveraging these technologies for maximum public benefit while simultaneously understanding how adversaries might exploit similar systems for asymmetric warfare," he said. "This approach enables agencies to continuously refine their AI workflows, building institutional knowledge that strengthens both defensive capabilities and offensive preparedness."

Clark also echoed Mulligan's point that AI in defense systems will both expedite daily workflows and teach the federal workforce about the technology itself. 

"It's critical that we use this technology to invest in our intelligence and defense ecosystems — this will both help them move faster to deal with a rapidly moving threat landscape, and it will also help them gain a better understanding of AI itself," he said.

---