SIEMENS

It's an eternal battle. Criminals conceive ways to hack into computer systems and send their creations through the Internet, while software developers at companies and public agencies scramble to block intrusions and protect users from damage. But as soon as one dangerous code is cracked, a new one springs up in the virtual world. Malicious software can trigger undesirable functions in an infected computer. Such programs usually operate in the background, unnoticed by computer users — but they can cause considerable damage, including manipulation or deletion of entire files, detrimental changes to a computer's online security software, and unauthorized gathering of data for marketing purposes or for spying on the user.
In 2005 the FBI conducted a study that included a survey of security experts at American companies, government agencies, and universities. The survey respondents reported an average loss of approximately $200,000 as a result of online crime. The most commonly cited causes of the financial damages were viruses and worms.
A particularly nasty form of attack is the computer worm. Unlike a virus, which attaches itself to a file that must be opened by the user in order to do its damage, a worm autonomously spreads through networks and attempts to burrow into other computers. To date, roughly 2,000 computer worms have been uncovered. Their codes are known, so their bit sequences can be "sniffed out" and barred by suitable security software, known as "sniffers." But new worms, different from those already known, can emerge every day. That's why we all need "digital watchmen" — electronic systems that can automatically identify these new threats.
With this in mind, specialists at the Learning Systems department of Siemens Corporate Technology (CT) are working on far-sighted solutions designed to permanently put worms out of business. Their weapon: development of a combination of neural networks with a quantum computer.

Less Wiggle Room for Worms. When it comes to neural networks, Siemens experts rely on processes that work in a way similar to those in the human brain. "In neural networks, all nodes are connected with one another. The information is found in the strength of the individual connections," explains Dr. Rudolf Sollacher, of the Learning Systems center at CT. "These networks can be used to recognize new patterns without having to be given specific rules, for example." Specialists call this adaptive learning.
For instance, with regard to interpreting handwriting, Christof Störmann, who works closely with Sollacher, explains that, "A neural network that is trained using the letter ‘a' written in many different ways will learn in time how an ‘a' is supposed to look and then will be able to recognize an ‘a' written in a style different from the ones it has already learned." It is this nearly magical ability of neural networks that the Siemens researchers are now taking advantage of in their efforts to hunt down new worms. They can use it to recognize potentially dangerous signatures, much like the way fingerprints are used to recognize criminals. However, if a worm is based on an entirely new process, a neural network will not catch it anymore than an FBI database will catch a first-time criminal.
This is why Störmann recommends taking the opposite approach. "A sniffer that rapidly recognize patterns reads entire data transfer processes. It allows everything to pass through that corresponds to normal business transactions. But the instant it detects something new and unexpected, it sounds an alarm. Then the task is to determine if the anomaly represents a threat," he says. This method is particularly suited for use in companies where the data being transferred internally is very well known.
Regardless of the anomaly-identification methodology, however, both require amazing speed. In fact, conventional computers trying to do the same thing would quickly be hindered by their physical limitations. With this in mind, Sollacher's team is planning to combine neural networks with a technology that is still in its infancy: quantum computing. "For complex tasks like pattern recognition, a quantum computer really has what it takes," say Sollacher.
Trained as a physicist, Sollacher (48) uses a special property of quantum systems for his work. As a rule quanta are not in a single, unambiguous state but rather are simultaneously in all possible states, "superposed" on one another. A conventional computer calculates with bits, which have a value of 0 or 1, but a bit in a quantum computer can be 0 and 1 at the same time due to superpositioning. In the quantum world this kind of bit has even been assigned its own official name: "qubit."

Exploiting a Universe of Possibilities. A system composed of two qubits can assume the following states: 00, 01, 10, and 11 — and all simultaneously. And the number of possible combinations rises rapidly. For instance, with 32 qubits there are four billion possibilities. The idea in quantum computing is therefore to exploit this multiplicity, with each calculation proceeding in all states at the same time. It is therefore easy to see why the achievement of such technology could result in a super-powerful parallel computer capable of working many times faster than conventional computers when recognizing patterns. And such a feature is a must if a sniffer is expected to check for malicious codes in real time in an avalanche of data that can easily amount to many gigabits per second.
There's one big problem with all of this, however: The kinds of quantum computers that would be needed to perform such operations don't exist. What's more, those quantum computers that have been developed to date can handle only a few qubits, and these prototypes are too complex and cumbersome for everyday use. That's why Sollacher and his colleagues decided to assess the advantages of running a neural network on a quantum computer by simulating how such a system would work.

Simulating a Quantum Network. The man who was given this complex simulation assignment is 29-year-old quantum computing specialist Rodion Neigovzen. "I started by transferring the spatial and chronological development of a neural network into the quantum world and developing the associated mathematical formulas," he reports. Neigovzen used a conventional computer to simulate this process, an original development for his doctoral dissertation under the supervision of Prof. Wilhelm Zwerger, the chair of Munich Technical University's Theoretical Physics department. "This simulation works with as many network nodes as you like," says Neigovzen, "but only if the computer has the required capacity."
Neigovzen's simulation of a neural quantum network functions beautifully in the virtual world. In fact, Sollacher's team at CT has already used the algorithm to predict how a real quantum computer running on a neural network would behave in detecting patterns. But do the results reflect reality? To determine the answer, Sollacher's team looked for a way to observe the results in an experiment.
Their search led them to Prof. Steffen Glaser and his colleagues at Munich Technical University's Chemistry department, where researchers had been working for years toward the realization of quantum computers and the ability to control qubits in theory and practice.
This encounter quickly led to the the establishment of a partnership between Siemens Corporate Technology and the Technical University — a partnership that soon produced an exciting breakthrough. In December 2007, in the basement of the Chemistry building, the two teams completed the world's first experiment creating a neural network which consisted of two bits running on a simple quantum computer.

Reality Check. The researchers who conducted the experiment, Neigovzen and the university's Dr. Jorge Neves, used a solution of water and sodium formate, whose molecules each have one carbon and one hydrogen atom. They poured the solution into a test tube, which they placed in a nuclear magnetic resonance spectrometer. NMR spectrometry is often used by chemists to perform structural analysis of biomolecules. However, the method can also be used to manipulate qubits.
The NMR principle is based on the fact that most atomic nuclei — and particularly hydrogen nuclei — behave like tiny bar magnets, spinning and tilting in a magnetic field. Thus, when the test solution is placed in a powerful magnetic field, its atomic nuclei arrange themselves along the lines of the field, as a result of their spin and magnetic moment. Then, using appropriate high-frequency pulses, the atomic nuclei are "perturbed" — the equivalent of entering information into the quantum computer. The nuclei start to rotate like tops around the lines of the magnetic field, giving off characteristic radiation, which can be measured. This step corresponds to reading out the desired data. In the case of the researchers' neural quantum computer, the measured signals agreed exactly with the values predicted by the simulation, confirming that Neigovzen's simulation of a quantum computer delivers correct results in practice.
As far as Prof. Glaser was concerned, the experiment constituted something like a contact with another world — the world of neural networks. "It's exciting to combine Siemens expertise in this area with our know-how in quantum computing," says Glaser.
And for Siemens, this collaboration could be the seed that grows into a neural quantum computer, one that can detect computer worms faster and more effectively than any system available today.
A prototype of such a computer is expected to be ready in one to two years. "But it will still be a few years before it all results in a product," says Sollacher.